Feb 2025 to Jul 2026
AI Engineer, TapHealth
AI Engineer from Apr 2025, previously AI Intern from Feb 2025, promoted after 2 months.
- Built and shipped an offline diabetes guidance SLM running fully on device on Snapdragon 680 hardware, in partnership with Eli Lilly. 380 MB, under 890ms latency at 25 tok/s, 98% accuracy on out of distribution refusal.
- Built a stateful 5 agent coaching system with LangGraph and Mem0/Hindsight memory across 25 tools. An async OCR pipeline over SSE WebSockets cut report processing from 40s to 15s.
- Cut LLM costs 54% by designing a centralized LiteLLM gateway routing 65+ agentic workflows across OpenAI, Anthropic, and open source models, with structured outputs via Instructor and full Langfuse tracing.
- Rebuilt an 812 line rules based nudge pipeline into a decoupled DSPy architecture (strategy classifier separated from content generator), serving 5K+ users.
- Started as AI Intern: built ARIMA, SARIMA, and LSTM forecasting models for patient engagement, and generated synthetic clinical data using NHANES and Q learning.
Jul 2024 to Oct 2024
Project Intern, Perception & Intelligence Lab, IIT Kanpur
Curated a 200K+ image dataset and optimized diffusion based inpainting models, achieving 11.72 dB PSNR and 0.41 SSIM.
Adversarial multi-agent fact-checker. A Prosecutor and Defender argue opposite sides of a claim over retrieved evidence, a deterministic verifier strikes any quote not found verbatim in the source, and a Judge rules. 96% quote grounding on a 32-claim eval. Ships as CLI, streaming web UI, and MCP server.
PythonLangGraphMCP
Spec-driven engine for AI-narrated motion-graphic reels. JSON spec → Gemini TTS → GSAP composition → MP4. Live demos at
@ada.max24.
PythonGSAP
A GPT built from first principles in pure PyTorch, no HuggingFace. Attention, RMSNorm, RoPE, SwiGLU, grouped-query attention and a from-scratch BPE tokenizer, each tested before the next was added. A same-seed ablation improved validation loss with 12% fewer parameters. Ships with a
live attention visualizer whose architecture diagram is generated from the model config, not hand-drawn.
PythonPyTorch
Due diligence for every GitHub repo. Know before you clone. Local, reproducible scoring.
JavaScript
On-device RAG for fast coding problem lookup.
PythonFastAPIDeepseekR1