Problem

Running several strategies across brokers by hand does not scale, and a single unverified cancel or flatten can flip a position and lose real money. The platform needed to execute reliably, know its true state at all times, and only promote strategy changes backed by evidence.

Outcome

The system runs unattended, reconciles itself against every connected broker, and every rule change is graded in shadow mode before it can trade.

What we built

• Signal ingestion from TradingView webhooks and direct broker market data
• A verified-action execution layer that confirms every cancel and flatten against the broker before proceeding
• Broker-to-ledger reconciliation with orphan-order detection and auto-heal
• Per-strategy daily risk caps and green-day give-back protection
• Shadow-mode strategy validation with weekly grading before anything goes live
• Regime gates and entry-quality filters derived from post-mortem analysis

Problem

Per-call API costs for market analysis were growing, and there was no honest way to tell whether the model added anything over a simple rule.

Outcome

A measurable path from POC to production, with the cost per analysis heading to zero. A POC that proves the model adds nothing over a baseline is treated as a successful result.

What we built

• Schema-enforced FastAPI service returning long / short / no-trade with entry, stop, target and confidence
• Ollama-served 7B–8B open-weights models with careful context engineering
• A replayable evaluation harness with rule-based baselines and strict no-lookahead checks
• Retrieval of similar historical setups to ground the model's reasoning
• A roadmap from proof-of-concept through RAG to LoRA fine-tuning

Problem

Holdings spread across countries, currencies and account types had no single source of truth, and spreadsheets could not keep multiple people in sync.

Outcome

One place to see net worth, allocation and movement across every entity and country.

What we built

• Multi-user application with per-entity portfolios
• US and Indian asset classes with currency handling
• SQLite with Prisma for a zero-ops, portable data layer
• A fast Vite front end for daily use

Problem

General-purpose speech models perform poorly on Malayalam. The team needed accurate transcripts at scale without manual re-typing.

Outcome

A repeatable pipeline with materially better Malayalam accuracy than the default choice.

What we built

• Batch pipeline against Sarvam's Indic speech models
• Side-by-side evaluation against Whisper on real recordings
• Reusable scripts for chunking, submission and assembly