Founder, AcoustiNode Systems · Performance bug hunter in open-source AI infrastructure
Embedded systems · AI · IoT
📄 IEEE paper published in IEEE Xplore (co-author, ICTEST 2026)
🏆 National Champion, Techletics 4.0 · 🏅 Samsung Solve for Tomorrow 2026, Top 40
📜 2 provisional patents
🧪 6 published research notes with DOIs
Featured in The Hindu BusinessLine · Moneycontrol · Malayala Manorama · Mathrubhumi · Samsung Newsroom
Many PDFs in Indian scripts look right on screen, but the text that search engines, screen readers and AI training pipelines read from them is garbled. I traced the problem, measured how big it is, and built the three tools that catch it, repair it and prove the repair, all without OCR.
| Step | Project | Key result | |
|---|---|---|---|
| ① Found | FinePDFs audit Hugging Face's FinePDFs AI training dataset, traced back to the PDFs |
~11B tokens of bad text, rough extrapolation (AI OCR invented ~4.8B in 7 Indian and Sri Lankan scripts; ~6B broken Persian filed under small languages); 94% of 791 sampled Indian government regional-language PDFs have no correct text (scans or garbled); in a controlled test, Word 365's Save as PDF garbled ~15 scripts (letter F1 97-99.5% for Edge vs 34-90% for Word, same text and font) | details ↓ |
| ② Detect | indic-pdf-check · pip install indic-pdf-checkIs this PDF's text broken? Answers in a second, on your own computer |
Word 365 PDFs with known-broken text in 14 scripts: 12 flagged | repo |
| ③ Recover | Unshape pip install unshapeExact text from broken PDFs, no OCR: 10 Indian scripts + Sinhala |
99.4% exact words on Word 365 PDFs; 98.5% on blind real PDFs from one publisher (PyMuPDF 72.4%) | details ↓ |
| ④ Prove | SeenRead pip install seenreadProves, word by word, whether extracted text is the text a human sees (also catches font-remap tricks and hidden text) |
Fresh blind tests: 99.98% of words exact, 0 harmful errors in the latest; beats Tesseract OCR (92-98%) in English, Hindi, Bengali, Tamil | details ↓ |
pip install indic-pdf-check unshape seenread # detect -> recover -> prove| Project | What it is | Key result | |
|---|---|---|---|
| exact-llm |
Bit-exact LLM inference: the same output bits on any machine, so anyone can re-check an AI provider's answer | Same bits on CPU and GPU from 0.5B to 7B, within 0.15% (0.5B/1.5B) and 0.41% (7B) of float perplexity; batch-1 decoding at 1.04–1.26× the speed of an equally optimised float16 version on a T4 | details ↓ |
| spectrace |
Predicting speculative-decoding gains without running the drafter | Matched real engines within 0.02–0.31% | details ↓ |
| resync-apply |
Lossless copy drafting for AI code editors | 1.29× fewer model passes than Snowflake's SuffixDecoding | details ↓ |
| prompt-injection-relocation | Stress test for AI-agent prompt-injection defenses | A gate that stopped 88% of attacks dropped to 28–62% when attacks moved channel | details ↓ |
I find code that gets slower (or hungrier for memory) much faster than it should as inputs grow, or that gives wrong answers where two implementations should agree, prove it, and send a fix with tests. AI agents help me search; every number below was re-measured by me, because most claims don't survive verification.
1 fix shipped by Hugging Face · 3 bugs fixed by maintainers (Meta FAISS, Hugging Face tokenizers, OmegaConf) · 12 PRs merged, into uvicorn (FastAPI's server), redis-py (the official Redis client, 5 PRs), nats.py (the NATS Python client), Hugging Face datasets, xarray (NumFOCUS, 2 PRs), Prefect and Black (PSF) · 9 more under review at uvicorn, Apache Airflow, ONNX, OGX, Meta (PyTorch, FAISS), Hugging Face, dask and CNCF Jaeger
| Project | Result | Status | |
|---|---|---|---|
| ✅ | Hugging Face Hubhf cache prune / hf cache rm |
79.8 s → 3.3 s 10,000-file cache |
Fixed in 36 h, crediting my suggested fix · issue #5100 · fix #5118 |
| ✅ | uvicorn (FastAPI's server)--reload startup |
120 s → 0.39 s 2,000 directories; users reported 2-min startups |
Merged by the maintainer · PR #3184 |
| ✅ | redis-py (official Redis client) sorted-set scores in the default mode |
Correctness fix scores came back with 17 digits ( 0.1 → 0.10000000000000001); on a real Redis server 74,109 of 120,618 scores were wrong, 116 after the fix |
Merged by the maintainer · PR #4427 |
| ✅ | redis-py (official Redis client) sorted-set scores with decode_responses=True |
Correctness fix scores came back as bytes where protocol 2 returns text ( zmpop, zdiff, zrank...); checked on a real Redis server and a 3-node cluster |
Merged by the maintainer · reported in issue #4428 at the maintainer's request · fixed by PR #4432 |
| ✅ | redis-py (official Redis client) MultiDB clients ( MultiDBClient and its pipeline) |
Crash fixslowlog_get(), acl_setuser() and other commands raised AttributeError on MultiDB clients, sync and asyncio; checked on a real Redis server |
Merged by the maintainer · reported in issue #4440 at the maintainer's request · fixed by PR #4442 |
| ✅ | redis-py (official Redis client) search through an asyncio pipeline ( pipeline().ft()) |
Crash fixsearch() and aliaslist() raised TypeError before anything reached the server; found while adding a test the maintainer asked for; checked on a real Redis server |
Merged by the maintainer · sent as a separate PR at the maintainer's request · PR #4491 |
| ✅ | redis-py (official Redis client) search commands in a pipeline ( info, profile, spellcheck, sugget...) |
Hang fix 8 search commands parsed the pipeline instead of returning it: info(), profile(), spellcheck() and config_get() never returned on a sync pipeline, and asyncio raised TypeError; checked on a real Redis server in 4 protocol modes |
Merged by the maintainer · sent as a separate PR at the maintainer's request · PR #4499 |
| ✅ | nats.py (NATS Python client, CNCF) JetStream pull fetch() after a server restart |
Hang fix fetch stayed stuck after a restart or reconnect (forever with no timeout); on a real server 0 of 20 messages arrived, 20 of 20 in 0.3 s after the fix |
Merged by the maintainer · PR #1065 · fixes issue #504, open since 2023 |
| ✅ | Hugging Face datasets loading datasets with many columns |
15–25 min → ~30 s real 20,531-column dataset, identical output |
Merged by the maintainer · reported in issue #8776 · fixed by PR #8777 |
| ✅ | xarray (NumFOCUS).str methods on text from pandas |
Crash fix 10 string methods ( split, pad, format...) crashed on object-dtype text, e.g. after DataFrame.to_xarray(); found by differential fuzzing |
Merged by the maintainer · fix with 10 tests · PR #11722 |
| ✅ | xarray (NumFOCUS).str.replace / .str.contains with case=False |
Correctness fixreplace(..., case=False, regex=False) silently ignored case=False, and contains crashed on object-dtype text |
Merged by the maintainer · fix with tests · PR #11747 |
| ✅ | Prefect (workflow orchestration) saving blocks that hold a list of credentials |
Silent data-corruption fix 18 of 20 saved entries came back wrong; 0 after the fix |
Merged by the maintainer · reported with a fix and tests in issue #23403 · fixed by PR #23404 |
| ✅ | Black (PSF, the Python formatter)# fmt: skip on an opening bracket |
Crash fix Black failed with "produced invalid code" since 26.10.0; bisected to the change that caused it |
Merged by the maintainer · fix with a test · PR #5538 |
| ✅ | OmegaConf (config library under Hydra)OmegaConf.missing_keys() |
Correctness fix keys containing . or [ ] were reported with wrong paths, so two missing keys could merge into one |
Reported by me, fixed by the maintainer the next day · issue #1423 · fix #1428 |
| ✅ | FAISS (Meta) exact k-NN search called inside a parallel region |
Silent wrong-results fix with few queries and a large index, search returned wrong neighbours (nearest distance 0 instead of 4.29); in Meta's own benchmark an IVF-Flat search found 0% of the correct neighbours before the fix, 54% after (the normal level for that setting) |
Reported by me, fixed by Meta · Meta's regression test cites my report · issue #5702 · fix #5824 · test |
| ✅ | Hugging Face tokenizers (v1) loading ai4bharat/IndicBERTv2-MLM-only |
Crash fix an Indian-language model (25k downloads a month) worked in 0.23.3 but failed on v1 with "keys must not contain NULL", because its vocabulary has NUL-byte tokens |
Reported by me, fixed by the maintainer the next day · issue #2660 · fix #2661 |
| 🐞 | uvicorn (FastAPI's server) new zttp HTTP engine |
Crash fixHEAD requests with non-ASCII URLs got no response; found by differential fuzzing |
Bug from the day before, reported with a fix · PR #3201 |
| 🌀 | Apache Airflow Graph view for large task groups |
10.3 s → 1.15 s two task groups of 800 tasks, identical output |
PR #74574 |
| 🧮 | ONNX (Linux Foundation AI) reference evaluator, ReduceSumSquare |
Correctness fix 32-bit integer input came back as 64-bit, against the spec; found by differential fuzzing |
Fix with tests · PR #8600 |
| 🧲 | OGX (formerly Meta's llama-stack) sqlite-vec vector search with a filter |
Silent wrong-results fix filtered searches returned too few results, often none (0/10 → 10/10) |
Reported with a fix and tests · issue #6778 · PR #6779 |
| 🧠 | sentence-transformers IR evaluator memory |
20.6 GB → 5.5 GB real NQ data, 32k queries |
PR #4148 |
| 🔎 | FAISS (Meta) binary k-NN, single query |
2.0–3.2× faster 2.02× on real binary embeddings |
Plus a nested-OpenMP correctness fix · PR #5701 · benchmarks |
| 🔥 | PyTorch (Meta)nn.Module.load_state_dict |
159.8 s → 2.8 s real forecasting model, 10,100 variables |
Root cause of a 5-year-old issue · analysis on #66787 |
| 🧩 | dask graph optimization for long pipelines |
32.2 s → 0.80 s 2,000-operation chain, identical results |
Fix for a 3-year-old open issue · comment on #9795 |
| 🧭 | Jaeger (CNCF) trace clock-skew adjuster |
Correctness fix follows-from spans wrongly shifted |
Issue open since 2017 · proposal on #448 |
"open", "proposed" and "posted" mean under review, not merged yet.
How I find them
- Scaling ladders: run an operation at sizes 1×, 2×, 4×, 8×… and measure time and peak memory. Anything growing faster than it should is a candidate.
- Differential fuzzing: send the same random inputs to two implementations that must agree (e.g. uvicorn's HTTP engines) and investigate every difference.
- Invariance checks: change settings that must not change the answer (batch size, chunk size, order, dtype) and diff the outputs.
- Floor check: compare real cost with the physical minimum (bytes moved, work done).
- Prove it on real data, check the fixed version gives identical results, run the project's tests, then report.
One agent once told me it had found a 700× speedup "with 100% confidence". The real number was 59–310×, and my own test harness was wrong twice before I trusted it.
In 2018 I watched floodwater swallow the roads around Chalakudy. Eight years later I'm still working on that problem.
AcoustiNode is a VHF radio telemetry network for dams and reservoirs. It moves water-level and flow data between upstream, dam-site and downstream points over a link that stays up when cellular and internet go down, which is exactly when release decisions matter most.
- Built the prototype myself: sensors, Raspberry Pi and Arduino edge nodes, radio encoding, Python
- Demonstrated at Peechi Dam to the KERI Director, Joint Director and Irrigation Department officials
Working with Government Engineering College Idukki on a system for managing a cascade of dams (several dams on the same river system).
- Co-inventor, provisional patent for the dam cascade management system
| 🥇 | 1st place, Techletics 4.0 national competition |
| 🏅 | Top 40 national semi-finalist, Samsung Solve for Tomorrow 2026 |
| 📜 | 2 provisional patents |
| 📄 | Co-author, IEEE paper: IoT-Based Resilient Monitoring Framework for Real-Time Safety of Interconnected Dams, ICTEST 2026 · IEEE Xplore |
| 📰 | Featured in Moneycontrol, The Hindu BusinessLine, EdexLive, Mathrubhumi, Manorama, Analytics India Magazine, Samsung Newsroom and more |
Float arithmetic gives different low-order bits on different hardware, so two honest AI servers return different numbers for the same request, and nobody can re-check what a provider computed. exact-llm defines the model as an integer program (int8 weights, integer activations, exact integer sums, integer tables), so every conforming implementation returns the same bits.
- Same output hashes on x86 CPUs and an NVIDIA T4 (float64 and int8 tensor cores), for Qwen2.5 0.5B, 1.5B and 7B
- Perplexity vs float32: +0.14% / −0.07% (0.5B / 1.5B, group-128 weights), +0.41% at 7B
- Batch-1 decoding on a T4, both sides compiled and CUDA-graphed in the same harness: 1.04× (0.5B) and 1.06–1.26× (1.5B) the speed of float16; exact 7B at 19.3 tok/s, where float16 7B does not fit in memory
- Built on exactness: a provider that quietly cuts one layer to 7 bits is caught on the first request; sampled text comes with receipts anyone can replay; one spot-check of a 7B answer takes about a second on a CPU
- Honest limits: prompt processing (prefill) is still 3–5× slower than float16; one GPU type; no comparison with vLLM / TensorRT-LLM yet. Builds on and credits concurrent work (tritify, samebit, bitexact-llm, int-llm, and others)
README · v0.4.0 release · Apache-2.0
PDFs can show people one text while giving machines another: a font that draws "1" where its table says "9", an edited text table, white or invisible text. AI pipelines, CV screeners, search engines and screen readers read the machine version. SeenRead identifies every glyph by its exact outline in a trusted font and accepts a text only if the shaping engine lays it out into the same glyphs - so for every word it proves SAME, DIFFERENT (with the visible text) or UNKNOWN, and lists text that machines read but humans cannot see. No OCR, no guessing; works for Indian scripts too.
- Two fresh pre-registered blind tests (120 new documents each, English, Hindi, Bengali, Tamil): 99.98% of words exact; the latest with 0 harmful errors; all hidden text found, no false alarms
- Same files with Tesseract OCR: 91.8-98.2% of words; an OCR-based lie detector caught 79-95% of lies with up to 4% false alarms
- Honest limits: synthetic documents so far; needs the trusted fonts; real-world PDFs are the next test
README · v0.1.0 release · AGPL-3.0 or commercial licence (open for talks)
FinePDFs audit: invented and broken text in an AI training dataset, government PDFs without correct text, and Word's PDF export 
FinePDFs (Hugging Face) is a 3-trillion-token dataset of text extracted from PDFs, used to train AI models. I audited its complex-script languages and traced every problem back to the original PDFs.
- Fabricated text: most documents in several scripts went through an AI OCR model; its stored text almost never matched an independent OCR of the same PDF in 7 scripts (Malayalam 18/18, Telugu 17/17, Kannada 17/17, Tamil 6/7, Sinhala 15/15, Odia 19/19, Punjabi 14/16), while Hindi, Bengali, Marathi, Nepali and Gujarati were fine. Checked by hand: a ministry pension scheme became "a platform for farmers in Kerala"; roughly 4.8 billion tokens if the sampled rates hold (rough extrapolation), and no signal in the dataset flags it
- Small languages get the garbage: the Southern Balochi (99%), Kashmiri (94%), Gilaki (73%) and Mesopotamian Arabic (69%) configs are mostly broken Persian or standard Arabic text, about 6.1 billion tokens (rough); big languages stay clean
- Government PDFs: 93.8% of 791 randomly sampled Indian government (gov.in / nic.in) PDFs in nine regional languages have no correct text (73% had no usable text layer, mostly scans; 21% garbled text layers, mostly Word-made in a checked subset; 6% clean), checked on 29 originals; no improvement between 2017-19 and 2023-25. A sample of what Common Crawl and FinePDFs kept
- A cause, a Word bug: with the same text and the same font, Microsoft Edge's PDF loses no letters (a few doubled vowel signs) while Word 365's Save as PDF loses and garbles many (letter F1 34-90%: Khmer, Myanmar, Malayalam, Hindi); Word's export keeps 7-65% of words readable in 10 Indian scripts and also garbles Divehi and Lao (about 15 scripts, over a billion speakers), while Arabic, Hebrew and Thai survive (measured on Word build 16.0.20430, 2026-10-17; build 16.0.20527 keeps these letters intact except Divehi, while the NVDA screen reader in Chrome still reads about 1 word in 3 wrong; details in finepdfs-audit)
- Recovery without OCR: reversing font shaping recovered 99.1% of words on held-out PDFs in 10 scripts, vs 86-91% for free OCR and 61% for the AI OCR the dataset used (research version; now released as Unshape, a few seconds per page on one CPU core)
Report · Code & results · Reported to Hugging Face · report CC BY 4.0, code Apache-2.0
Copy-paste, search, screen readers and AI pipelines often get garbled text from Hindi and other Indian-language PDFs. Unshape recovers the text from the glyphs themselves: it searches for the text whose font shaping reproduces the PDF's glyphs and keeps a stretch only if it re-renders exactly. When the embedded font has lost its shaping tables (Word exports), it matches the glyphs to the original font the user supplies.
- Word 365 PDFs in 10 Indian scripts + Sinhala: 99.4% exact words (first frozen run, 7123/7164)
- Blind real PDFs from one publisher: 98.5% (PyMuPDF 72.4%); every blind test is listed in the README
- Honest limits: needs the original font; Mangal (the most common Word Hindi font) not tested yet; no scans
README · v0.2.0 release · AGPL-3.0 or commercial licence (open for talks)
AI code editors apply an edit by having a model rewrite the whole file, most of it copied. I built a lossless "copy-guesser" that lets the model confirm long copied spans in one pass, and compared it with Snowflake's SuffixDecoding, the strongest open method (built into vLLM).
- 1.29× fewer model passes than SuffixDecoding on edits never used for tuning (95% CI 1.22–1.37), output identical to normal decoding
- 24–37× faster than normal decoding on a free T4 GPU; against SuffixDecoding the wall-clock was a tie on that GPU, and the report explains why
- A laptop simulator that predicted the GPU pass counts exactly in three runs, plus 7 ideas that failed, with numbers
Report (PDF) · Code · Apache-2.0
Speculative decoding speeds up a large model by letting a small "drafter" guess ahead; choosing the right drafter usually means running each one. I tested an offline way to predict a drafter's gain from outputs the model has already produced (the estimator was published first by Zhao & Cai, arXiv:2610.10411; I derived it independently and credit them), checked it against real engines, and applied it to choosing drafters during RL training.
- Predictions matched real engines: 27.36 vs 27.28 passes (Hugging Face, sampling); 931 = 931 (greedy); the engine's own probabilities matched within 0.02–0.31%
- Picked the best of 3 drafters correctly for 2 models, within 0.2–1.2%, about 140× cheaper than measuring (measured 144×); held up across a real RL-trained model
- An exact replay check for vLLM's speculative-decoding counters independently caught two known metric bugs, one of them ~10× off
prompt-injection-relocation: a stress test that moves logged prompt-injection attacks into other channels an agent reads (email, files, calendar, web).
- A defense gate that stopped 88% of attacks where they were originally placed dropped to 28–62% in 10 of 16 channel types
- A redesigned gate held 91–100% in 14 of 16 (weak spot: delegated sources)
- Offline replay, reproducible in about 5 minutes
Open to internships and collaborations in AI infrastructure, performance engineering and embedded/IoT,
and to being told I'm wrong about something.
LinkedIn · Google Scholar · ORCID · Student: Computer Science with Physics, Chemistry and Maths


