I'm an AI Engineer who builds and ships full-stack production AI systems — solo, end-to-end. In the last year I've shipped a RAG learning platform, a voice dictation desktop app, a real-time automotive voice AI agent, and a multi-agent VS Code extension. I don't just prompt; I build the full stack around the AI layer.

Background: 3 years enterprise security automation at Wipro → MS Business Analytics at UC Davis (practicum at The Best Notary) → solo production AI product development.

What I work with daily: Claude Code, OpenAI Codex, FastAPI, Next.js, LangChain, Qdrant, LiveKit, Electron.

AI-powered personalized learning: ingest documents → extract concepts → adaptive quizzes → Bayesian mastery tracking

The problem: Students have no efficient way to turn dense study materials into active recall. AutoCoach automates the full study loop.

Architecture:

PDF/PPTX upload
    → LangChain chunking (1000 tokens / 200 overlap)
    → OpenAI batch embeddings → Qdrant vector store
    → Kimi K2.5 concept extraction (Pareto-corrected for score-inflation bias)
    → RAG-grounded quiz generation (document-scoped retrieval)
    → Three-tier answer evaluation: Kimi K2.5 → GPT-4o-mini → length heuristic
    → Bayesian-smoothed + EMA mastery update
    → Spaced repetition queue (resurface if mastery < 75% AND last tested > 2 days)

Key decisions:

• Pareto post-processing on concept extraction — fixed systematic LLM score-inflation bias; top 20% classified as "core" concepts driving the primary KPI
• Bayesian mastery formula: smoothed = (correct+1)/(correct+tested+2), new_score = 0.85*old + 0.15*smoothed — prevents overconfident early mastery; caps at 95% until ≥5 tests AND ≥90% accuracy
• Zero hard-failure paths — three-tier fallback means no answer evaluation ever crashes
• Optimistic locking on XP/credit writes — 5-retry pattern for concurrent quota safety

Cross-platform Electron app: floating bubble → Groq Whisper → Claude formatting → keyboard injection into any app

The problem: Voice-to-text tools are either cloud-dependent, context-unaware, or inject trash formatting. Voxi is local-first, model-composable, and clipboard-aware.

Architecture:

AudioWorklet PCM pipeline (16kHz mono, 250ms buffering)
    → Groq Whisper-Large-V3-Turbo (transcription)
    → Claude Haiku (optional formatting — cost-sensitive, high-frequency)
    → nut-js keyboard injection → focused app

"Hey Voxi" command mode:
    speech → intent extraction → clipboard read
    → Claude Sonnet with task-specific prompts (summarize / translate / debug / rewrite)
    → inject — zero UI interaction

Key decisions:

• Multi-model cost split: Claude Haiku for every dictation (fast, cheap), Claude Sonnet only for complex clipboard commands (slower, more capable) — deliberate latency/cost tradeoff
• AES-256-CBC key storage — API keys encrypted using machine-id-derived key, IV prepended in ciphertext
• Three-tsconfig build system — solved hard Electron ESM/CJS context boundary bug (renderer vs. preload vs. main)
• SQLite correction learning loop — stores raw→corrected pairs per app for automatic personalization without model fine-tuning
• Shipped as macOS universal DMG (arm64 + x64) + Windows NSIS installer

Ask a car question out loud, get a spoken answer in real time — specs, pricing, safety ratings, comparisons

Architecture:

LiveKit AgentSession
    → Deepgram Nova-3 STT (streaming, 25ms endpointing)
    → Cerebras Llama 3.1 8B (LLM)
    → Three-layer knowledge backend:
        (1) ChromaDB RAG + MiniLM-L6-v2 embeddings
        (2) SQLite cache (24h TTL + background preloading of top-10 cars)
        (3) asyncio.gather fan-out → NHTSA + fueleconomy.gov + CarQueryAPI + DuckDuckGo
    → Custom Smallest.ai TTS plugin (raw int16 PCM via AudioEmitter)
    → React Recharts latency HUD (STT + EOU + LLM TTFT + TTS, live)

Key decisions:

• Three-stage entity resolution: exact alias lookup (40+ entries: "miata"→Mazda MX-5) → fuzzy SequenceMatcher (≥0.55) → live NHTSA makes-list scan
• Function-call bleed guard — custom TTS wrapper drops text starting with <function= before synthesis (discovered through real LLM behavior in production)
• Provider-swap abstraction — iterated across 3 LLM providers (Groq → Cerebras → GPT) due to rate limits and tool-call format issues; clean interface kept swap cost minimal
• Per-turn latency instrumentation published over LiveKit data channel → live HUD

Forked from VS Code Marketplace · Shipped the OpenCode SSE integration · Visual pixel-art office tracking concurrent Claude Code + OpenCode agents

My contribution: The original Pixel Agents extension visualizes Claude Code agents as pixel-art characters in an office world. I forked it and shipped the OpenCode integration — a new SSE tracking architecture that enables the extension to monitor OpenCode agents via HTTP event streaming alongside Claude Code's JSONL file-watching, unified behind a single message abstraction so the webview is completely framework-agnostic.

The hard problem — multi-agent SSE routing:

Problem: OpenCode runs multiple concurrent sessions. Naive "mostRecentAgentId" binding
misroutes events when N agents launch simultaneously.

Solution:
  1. Pre-launch snapshot: capture existing session IDs before terminal starts
  2. Poll /session endpoint until a NEW session ID appears → bind that ID to the terminal
  3. sessionIdToAgentId dispatch table routes all subsequent SSE events correctly
  → Works correctly with 10+ simultaneous agents

Other technical highlights:

• Triple-layer file watching (OS fs.watch + fs.watchFile stat polling + manual setInterval) with partial-line buffering — prevents JSON parse errors mid-write on append-only JSONL
• Canvas 2D game engine outside React — OfficeState imperative class holds all game state, 60fps requestAnimationFrame loop runs without React reconciler diffs — smooth at 10+ agents in a sandboxed WebviewViewProvider
• Inference-based state detection — dual idle signals: system/turn_duration records + 5s silence timer for text-only turns; permission bubble inferred via 7s non-progress timer
• Claude Vision in asset pipeline — @anthropic-ai/sdk auto-generates sprite metadata from purchased tileset PNGs, feeding a 7-stage import pipeline to furniture-catalog.json

Real-time ELT: GCP Pub/Sub → BigQuery → dbt staging/intermediate/mart → liquidity runway analytics

Designed a streaming ingestion pipeline simulating real-time corporate treasury cash flows (receivables/payables). Built a multi-layer dbt warehouse in BigQuery with projected balance and liquidity runway analytics — automated end-to-end with no manual refresh steps.

End-to-end ML pipeline with SHAP explainability — black-box to business insight

Logistic Regression + Random Forest with GridSearchCV tuning, class-imbalance correction, and a SHAP pipeline that surfaces top churn drivers for non-technical stakeholders. The explainability layer was the whole point.

Multi-label chest pathology classification on NIH dataset with Grad-CAM visual interpretability

ResNet CNN predicting Atelectasis, Effusion, and Infiltration simultaneously from NIH Chest X-rays. Full evaluation pipeline (Precision/Recall/F1/ROC-AUC per condition) with Grad-CAM to highlight regions influencing predictions — interpretability matters when the stakes are clinical.

🎓 M.S. Business Analytics — University of California, Davis · 2023–2025 🎓 B.E. Computer Science — India · 2017–2021