A decade of enterprise systems. Now production AI, end to end.

01

Dialoft

AI voice SDR - outbound sales & qualification

An autonomous voice Sales Development Representative. One LangGraph conversation brain, exposed behind an OpenAI-compatible endpoint, that both Vapi (phone) and LiveKit (browser) call - so no logic is duplicated. It runs discovery and BANT qualification, handles objections from a knowledge base, books meetings, and enforces compliance as first-class steps.

• One brain, many channels: a LangGraph state machine (8 nodes, 7 state-derived stages) behind an OpenAI-compatible /v1 endpoint shared by Vapi and a LiveKit worker
• Deepgram Nova/Aura STT-TTS on both channels; Claude via langchain-anthropic as the reasoning model
• BANT slot-filling with a 0-100 lead score (HOT/WARM/COLD) recomputed each turn; 8 LangChain tools (CRM, calendar, KB, DNC, scoring)

02

Virtual Buddy

AI-employee platform for creators

A web platform where social-media creators hire a named, persistent AI teammate that tracks their socials, catches trends early, runs their calendar, and works as a conversational partner. Two paid tiers, live in production.

• Streaming Claude agent over tool-calling with tier-based model routing across Haiku, Sonnet, and Opus; tier and persona derived server-side
• Multi-tenant Postgres (14-table Drizzle schema on Neon) with row-level security via a non-bypass app_user role, verified end to end
• Vercel cron jobs (metric refresh + Haiku briefings) with per-run job tracking, failure webhooks, and a stale-job 503 health endpoint

03

VYUHA

Multi-agent trading desk for Indian markets

A configurable multi-agent AI trading desk for NSE/BSE equities and index options, structured like a real trading firm: specialist agents sense the market, a research floor argues the view, a hard risk gate sizes and vetoes, and an execution desk carries approved plans to close - all passing one typed Trade Ticket. Runs fully offline on synthetic data as a paper engine.

• 46 agents across ten layers - sensing, analysis, research debate, strategy, risk & governance, execution, reference data, middle office, ops, and a separate index-options desk
• One typed Trade Ticket is the contract between layers, advancing SCANNED → ANALYZED → RESEARCHED → PLANNED → SIZED → RISK_OK → OPEN → CLOSED, which makes every session auditable and replayable
• Middle/back office that most strategy code skips: STT/GST/stamp cost model, SPAN + exposure margin, parametric VaR and sector heat, square-root market-impact slippage, reconciliation, attribution, and a SEBI-style audit trail

04

Sentinel

AI risk coach for Indian options traders

An AI that watches live F&O positions and the trader's own behavior, and stops the mistakes that wipe out retail traders - over-leverage, revenge trading, averaging into losers, breaking your own rules. Read-only and single-user by design: it coaches on risk and behavior, never on what to buy, which keeps it outside SEBI's retail-algo framework.

• Broker-agnostic adapter layer (Groww REST, Zerodha Kite SDK, mock) normalizing to one Position/Margin type - adding a broker changes nothing downstream
• Deterministic rules engine for hard limits (day loss, margin utilization, concentration, position count) - pure code, instant and free; the LLM only does the fuzzy work
• Claude reads the SQLite trade timeline to flag behavioral patterns with evidence, plus a daily debrief graded A-F on process rather than P&L; on-demand only, never on a timer

05

AlphaPilot

Regime-adaptive trading agent

Backtests a library of systematic strategies, detects the current market regime, and deploys the best strategy for live conditions - sizing every position under strict risk limits and executing through a safety-gated broker. It does not bet on one strategy: it ranks by recent risk-adjusted performance and goes flat when nothing is working.

• Nine strategies across trend, reversion, and breakout families, scored by a vectorized look-ahead-safe backtester (signals shifted one bar, costs charged on turnover)
• Regime detection on two axes - trend/range and volatility - from a 2-of-3 composite of ADX, Kaufman Efficiency Ratio, and Choppiness, stabilized with hysteresis
• Walk-forward selection ranks on past returns only, holds through a minimum period with a switch margin to avoid whipsaw, and goes flat if nothing clears the bar