Case studies
Four real projects, with the numbers they produced. Every figure below was measured after delivery, not projected before it.
01SaaS Startup
Spending ₹85K/month on AI API calls with no cost controls or optimization.
Implemented smart caching, model tiering (GPT-4 for complex, GPT-3.5 for simple), request batching, and prompt optimization.
02Algo Trading Firm
Manual backtesting taking days per strategy. No systematic way to evaluate, compare, or iterate on trading algorithms.
Built a custom backtesting pipeline with automated data ingestion, strategy parameterization, walk-forward analysis, and visual performance reports.
03Digital Agency
Team spending 60+ hours/week on repetitive tasks — data entry, report generation, client communications.
Built a comprehensive automation suite covering CRM sync, invoice generation, weekly reporting, and Slack notifications. All custom-built, zero per-task pricing.
04Fintech Startup
Manual client onboarding taking 3 days per client. Document verification, KYC checks, and account setup all done by hand.
Built a multi-agent AI system using MCP protocol — one agent handles document extraction, another runs verification, a third creates accounts and sends personalized welcome flows.
05Crypto derivatives analytics
Backtesting 40 symbols took 300 seconds per run — too slow to iterate on strategy ideas, and the grid optimizer multiplied that pain by thousands of runs.
Re-engineered the engine over four versions: 2.3GB of raw tick data pre-aggregated to 1-minute Parquet, a numpy price grid for 40 symbols × 525K minutes, and the hot loops rebuilt after profiling named the real bottlenecks.
Strategy basket admin — btadmin.architmittal.in ↗06Crypto signal advisory
Picking strategies by eyeballing a handful of charts — no way to know if a "winner" was real edge or curve-fit luck.
Built an optimization pipeline that swept 64,320 parameter combinations in ~90 minutes, then a walk-forward filter: five years in-sample, five months out-of-sample — only strategies whose out-of-sample results reproduce the in-sample ones go live.
The live platform — btfull.architmittal.in ↗07Proprietary trading account
An account that felt like it was losing to the market. 178 real positions over six months said otherwise.
Forensic audit of every fill: fees of ₹37,966 exceeded the net trading loss itself; low-conviction days alone cost ₹36,408; realised payoff was 1.78:1 against a believed 1:10. Rebuilt sizing (5% → 1% probes with halt tiers) and moved venues to halve the round-trip cost.
08Private trading client
Wanted to capture exchange funding-rate windows 24/7 — without handing his API keys or capital to a third-party platform.
Delivered a funding-rate capture bot deployed on the client’s own VPS with a password-protected dashboard: he sets notional, leverage and thresholds himself. Isolated release branch so his version never breaks when the core evolves. ~77% win rate in our testing of the capture mode.
Engineering results from real systems — figures from project records. Nothing here is trading or investment advice.
One call, one process mapped for automation. The plan is yours either way.