Projects

Engineering notes on systems I built or contributed to — problem, architecture, decisions, and measured impact where I have it.

At 10xScale.ai

Domain-specialized medical agent

Private
Role owner Stack AgentFlow · FastAPI · SSE · Qdrant

Problem

Need a conversational medical agent that follows skill-specific playbooks and takes real actions — not open-ended chat.

What I built

  • StateGraph + ToolNode agent with per-skill playbooks loaded at runtime (triage, booking, medication guidance).
  • FastAPI service with SSE-streamed responses; checkpointer for short-term context; Qdrant for cross-session memory.
  • Transactional tools: doctor availability, appointment booking, structured DB queries.

Decision

Runtime playbooks over a monolithic prompt — skills stay swappable without redeploying the whole graph.

Schematic illustration of an edge voice assistant

Voice-enabled edge medical kiosk

Private
Role end-to-end owner Stack Raspberry Pi 5 · PySide6 · FastAPI · Twilio

Problem

Kiosk-side voice assistant that must feel responsive in Indic languages while staying wired to the same agent backend.

What I built

  • Async VAD → STT → Agent → TTS pipeline on Pi 5 with per-turn language auto-detection across 11 Indic languages.
  • Integrated backend APIs, agent service, web app, and desktop app into one production path.
  • WhatsApp via Twilio: inbound PDF/image extraction with Gemini Vision; outbound consultation summaries.
~70% barge-in latency ↓ (1000→300 ms) 11 Indic languages

Decision

Flush OS audio buffer on stop-TTS and chunk TTS at the first sentence — barge-in felt broken until both landed.

Enterprise AI interview platform

Private
Role agent + voice + eval Stack MCP · Electron · React · JSON-RPC

Problem

Voice interviewer that can run CV, DSA, system-design, and behavioral rounds with persistent state and automated coding evaluation.

What I built

  • Agent logic for dynamic questions, round progression, evaluation flow, MCP question bank.
  • Backend APIs to execute coding snippets against tests; transcript + per-round aggregation into a final report.
  • Real-time voice loop with VAD endpointing, TTS pre-warming, token streaming, barge-in; Electron sidecar integration.
sub-700 ms voice turn latency
Illustration of MCP bridging a database and tools

MCP server & database-aware agent

Private
Stack FastMCP · PostgreSQL · Typesense

Problem

Agents hallucinating facts about structured course/job data instead of querying the source of truth.

What I built

  • Custom MCP server exposing NL→SQL and Typesense keyword + embedding search, behind auth middleware.
  • Wired into a multi-agent database node that discovers tools, plans calls, and grounds answers in real data.

Impact

Factual questions stop depending on model memory — tools become the retrieval path.

Open source

Illustration of an agent orchestration graph

AgentFlow

Contribution memory + adapter

What I shipped

  • Qdrant-backed long-term memory module for cross-session agent context.
  • Hardened Google GenAI adapter against empty / None-content edge cases with regression tests.
↗ PR #17 ↗ PR #19
  • Python
  • Qdrant
  • AgentFlow

Personal & coursework

Illustration of a research agent searching a knowledge graph

Deep Research Agent

Context Hugging Face MCP Hackathon Stack Python · MCP · LangChain

What I built

Autonomous web research agent that researches topics, synthesizes sources, and writes comprehensive answers via MCP tooling.

Minimal bar chart illustration for churn modeling

Customer churn prediction

Stack Scikit-learn · SHAP · SMOTE

What I built

End-to-end telecom churn pipeline: class-imbalance handling, Random Forest tuning, SHAP interpretability, and threshold optimization against business objectives.

CIFAR-100 high-confidence classification

Coursework
When Sep ’24 – Nov ’24 Stack PyTorch · WideResNet · SAM

What I built

  • WideResNet (WRN-28-10) with Sharpness-Aware Minimization for generalization.
  • Temperature scaling calibration to improve reliability on high-confidence outputs.