Ranjan Shitole

Allocations · Dubai, UAE

• Built the company's AI assistant across backend and frontend — streaming SSE chat on Cerebras GLM-4.7 with Model Context Protocol tool calling over live fund, investor and banking data.
• Engineered its reliability layer: intent-based tool routing, token-budget pruning for long sessions, MCP reconnection, and server-side pagination keeping summaries consistent with tables.
• Authored the safety architecture — mandatory live-fetch over cached history, anti-hallucination rules, and reasoning-block suppression preventing system-prompt disclosure.
• Built three MCP servers with a full OAuth 2.0 provider (email OTP + TOTP 2FA), ES256 JWT validation, opaque token registry, headless auth, and 40+ tools.
• Built a gateway orchestrating 10 AI agents with classifier-based routing, multi-agent synthesis and SSE progress streaming.
• Built AllocationsX, the firm's FINRA-regulated pre-IPO secondaries trading platform — NestJS + Supabase API and Next.js admin console across 24 domain modules.
• Architected the legal document generation platform — six agreement types from structured deal data via Puppeteer HTML-to-PDF, with DocuSeal e-signature webhooks and SHA-256 audit trails.
• Led security hardening across 8 repositories and fixed a multi-gigabyte production memory leak traced to unbounded session state.

Baven (GigaBrain) · Pune, India

• Built AI trading agents executing live on-chain across Hyperliquid perps, Lighter (zkSync L2) and Polymarket — a 30-tool LLM interface with execution-safety guardrails on every order.
• Engineered the wallet layer: factory-deployed smart-contract wallets on Base, Privy delegated signing, EIP-712 typed-data signatures, BIP-32/39 derivation and HPKE key custody.
• Shipped x402 HTTP-native USDC micropayments on Base — pay-per-request API access enabling autonomous agent-to-agent commerce without accounts.
• Architected a 6,300-line async technical analysis engine — 25+ indicators, market-regime classification, pattern and divergence detection, and volatility-based risk sizing (~70% faster).
• Built two market-data SDKs from scratch: CoinGlass (284 typed models, ~120 endpoints) and DeFiLlama (12 domains).
• Designed multi-provider LLM routing with bring-your-own-model across OpenAI, Anthropic, Gemini, xAI and OpenRouter.
• Deployed 20+ Python microservices on AWS ECS Fargate with per-agent IAM isolation and SQLite + Litestream S3 replication for crash recovery.

Soul AI · Remote

• Reviewed and rated model outputs for RLHF training data across reasoning and code tasks.

Onelab Ventures · Hinjewadi, Pune

• Built and deployed an AI calling agent processing 320+ interviews at 98%+ transcription accuracy.
• Architected an async backend (15+ REST APIs, WebSockets) handling 70 concurrent calls with horizontal scaling — about 65% lower latency.
• Implemented a Transformer-based text generator with sub-second inference using PyTorch.

Ventory · Pune

• Developed backend features and extended API functionality in Python.

MaticAlgos · Pune

• Optimized real-time tick-by-tick data pipelines to run 3.2× faster, cutting latency roughly 69%.
• Developed 45+ technical indicators at >95% accuracy and a Black-Scholes options engine with full Greeks.
• Integrated 4 broker APIs (5Paisa, Kotak, Upstox, DHAN) over REST and WebSocket with OAuth2/JWT auth.
• Stood up a Loki–Prometheus–Grafana observability stack for live monitoring.