Work

Python · Multi-strategy decision engine · Live venue execution · Staged promotion

A decision system is only as good as the discipline around what it is allowed to do. Built a multi-strategy execution layer where every candidate strategy runs the same lifecycle — shadow, then paper, then live — and is measured against pre-registered criteria it declared before seeing results. Most candidates never graduate: over 60 were evaluated, 4 promoted, and more than 90% killed by their own test batteries. Said no to promoting on a good backtest — a strategy earns capital by surviving falsification, not by looking good in-sample.

Agent infrastructure · Falsification test batteries · Non-custodial copy service

Most of this industry markets performance and hides method. Built Glasshouse on the inverse: the product is verifiability. Systematic strategies are developed under falsification discipline and killed by their own test batteries before they touch capital, and the copy-service model is non-custodial — clients keep custody of their funds and compensation is performance-share only. The whole operation runs on the agent infrastructure below: research agents, monitoring loops, promotion gates. Said no to performance marketing — the pitch is the audit trail, not the returns.

Astro · Claude API (claude-opus-5) · Tool use · Prompt caching · SSE streaming · Vercel

Conversational AI clone on pavan.blog. Designed the system prompt intent-first: it works out who the visitor is (recruiter, collaborator, exploring) before answering, then pulls only the knowledge that fits that need through tool calls, and hands genuine leads to me. Every conversation logs its token cost (input, cache, output) so cost per conversation is measured, not guessed; prompt caching keeps the large stable prefix cheap. Said no to RAG — the corpus is small enough that on-demand knowledge sections are cheaper and simpler than a vector store.

Python · Flask · systemd · Webhook signal routing · Telegram alerts

Autonomous systems that move money need a guarded layer between the decision engine and execution — otherwise a model bug becomes a wallet bug. Built and operate a Python orchestration service routing signals from upstream decision systems through a risk guardian (drawdown-kill, per-strategy loss caps, conflict and duplicate guards, two-stage entry, tested kill switch) and staged-promotion gates (shadow → paper → live). Telegram alerts on every entry, close, and error. Said no to hooking every upstream system immediately — only ones that pass the pre-production gate are enabled. Public pattern extracted to GitHub; live dashboard and production adapters remain private.

Python · Flask · cron · Plug-in stream registry

Once you scale beyond two production streams, ad-hoc promotion decisions become the bottleneck — and the source of every avoidable incident. Built a Lab Framework where each candidate stream registers its own gate criteria (statistical thresholds, capital limits, error tolerances), and a nightly review cron measures every stream against its criteria. Two endpoints surface the state of the world: /api/live-readiness reports which streams have passed all gates, /api/risk-status reports which need attention. Said no to manual promotion overrides — every promotion is gate-driven and audit-logged.

Python · Pandas · DLD open-data pipeline · Eval harness · Vercel

Dubai real-estate decisions at Dash Capital were being made against scattered DLD exports — slow and hard to re-run. Started as a toolkit that loads and normalises DLD transactions; now a live Q&A over real 2026 residential sales. The hard product call was what it must NOT answer: the DLD feed only serves the current year, so year-on-year is refused rather than estimated, modelled yields are labelled as modelled, and a question it cannot answer from the data gets a refusal with answerable alternatives — never a confident guess. Evaluated on a question set written by a separate model that never saw the code: 85% (34/40) at first contact. Live questions feed a human-gated learning loop — failures become eval candidates, a person approves the label. Said no to an LLM-only answer path: a rules planner answers at $0 per question, and the LLM planner is benchmarked against it on the same evals before it earns a place.

Claude API · WhatsApp Business · Booking & CRM workflow automation

A UAE field-services business was losing leads to slow replies and spending hours daily on inquiry handling. Built (as a paid engagement) a WhatsApp-native AI assistant that answers inquiries, qualifies leads, quotes and books jobs 24/7 — with human-in-the-loop handoff for edge cases and pricing exceptions. Response time went from hours to seconds, and the owner reclaimed hours of admin every week. Said no to a custom app — met customers on the channel they already use.

Python · Anthropic Claude API · Cron-friendly

Content research is the slowest, most expensive part of running a niche newsletter when done by hand. Built two Claude-powered agents that turn a Monday morning's research into a 2-minute cron job: one agent surfaces trending topics, pain points, and regulatory updates (VARA, UAE Central Bank); the other runs a YouTube content-strategy brief with hook titles and content gaps. Said no to RAG, scraping, and vector DBs — a single structured prompt is enough for weekly cadence. Keep the complexity budget for downstream production.