one page, no filler

Sagar Institute of Science & Technology (SISTec), Bhopal

Bhopal (Remote). Building an AI operations platform for a US business client: a live operations dashboard and an owner-facing web app on Next.js, Supabase and the Claude API.

Remote. Building a 4-agent "chief of staff" with shared memory and a morning voice briefing (Vapi, ElevenLabs); RAG over sponsor files, with human approval before anything publishes.

Python · FastAPI · Qdrant · VoxCPM · Docker. Built a voice RAG assistant: Groq Whisper speech-to-text, Llama 3.2 over Qdrant for grounded answers, and VoxCPM TTS fine-tuned with a custom LoRA voice adapter, driving a lip-synced Three.js avatar. Split inference into 3 Docker services so a TTS crash cannot take down the API; deployed on a hardened VPS (key-only SSH, Fail2Ban).

Python · FastAPI · Instructor · SQLGlot. Built a Slack agent that turns plain-English business questions into read-only SQL and posts tables, charts and cohort grids, using one structured LLM call (intent, SQL, chart, confidence). Replaced regex blocklists with SQLGlot AST guardrails (read-only, table whitelist, alias-aware columns) and a retry loop that feeds validator errors back to the planner.

TypeScript · WebGPU · Three.js · CI. Implemented a lattice Boltzmann fluid solver in WGSL compute shaders, with GPU voxelization of any uploaded 3D model and FP16 storage to halve per-cell memory traffic. Built a validation suite that runs headless in CI against published references (cylinder Strouhal number 0.173 vs 0.166, sphere drag, Ahmed car body); interpolated walls more than halved sedan drag error (2.05 to 0.91).

Python · FastAPI · Ollama · Qdrant · SSE. Built a RAG backend that grounds answers in retrieved context and streams tokens over SSE, with consent-based tool calls and backend-enforced safety rules. Cut end-to-end latency from about 20 s to 6 s by bounding prompt size, gating retrieval and limiting memory injection.

Python · Whisper · spaCy. Built a speech-to-text tool that detects spoken topics, pulls related names into Whisper's prompt and re-decodes, so rare names come out right; ships a CLI and an Earnings-21 benchmark harness.

Merged PR #1134. Replaced fixed sleep-timer presets with a custom 5 min to 12 h timer; merged in v1.8.2, shipped to users in v1.9.

Python · TypeScript/JavaScript · C++ · SQL · WGSL

LLMs · RAG · Qdrant · embeddings · function calling · agents · prompt engineering · Claude/OpenAI/Gemini APIs · Ollama · Whisper · LoRA

FastAPI · NestJS · Next.js · Supabase · REST · SSE · webhooks · OAuth · PostgreSQL · Redis · Pandas

Git · Docker/Compose · Linux · Nginx · GitHub Actions (CI/CD) · Vercel · VPS hardening · FFmpeg · WebGPU

Data structures and algorithms · OOP · DBMS · Operating Systems