Taha Yasin BhattiEmail

Selected systems · 2023–26

Work

Production ML comes first here because that is where most systems become honest. Public energy and climate work follows; it shows the same engineering habits under different constraints.

Production ML for international teams

2023–26

Lead architecture and principal engineering

Agents and privacy controls across three operating systems

A multi-agent order-and-fulfilment assistant for a US retailer, taken from prototype to production against live order systems. I designed structured tool calls, permission boundaries and regression coverage, then turned the implementation into a reusable template for a second brand in the same group.

The same period included a package-recommendation agent for a telecom operator in the MENA region and a two-layer PII detection pipeline for a Canadian property group. Different sectors, same practical question: what must the model be allowed to see, decide and change?

  • 3 system shapes
  • 2 retail brands
  • 2-layer PII control

Agent orchestration · Structured tools · Retrieval · Evaluation · PII detection

system trace2023–26
01Request
02Orchestrator
03Policy
04Tools
05Trace
Disclosure-safe agent orchestration diagram

Energy and public infrastructure

2025–26

Lead architect

Multi-agent grid and EV planning

The planning surface combines five layers that usually live in different tools: distribution-grid data, traffic patterns, geospatial analysis, demand forecasts and hosting-capacity estimates. A FastAPI service coordinates the analysis. The Next.js and deck.gl front end handles city-scale polygon and vector-tile rendering; Docker Compose keeps the working system reproducible across environments.

The work was presented at the Asian Development Bank’s Asia Clean Energy Forum 2025 as an emerging-market clean-mobility case. I also presented EV-load and grid-integration modelling to senior planning leadership at a provincial distribution utility, and hosting-capacity analysis to the leadership of a national grid organisation. The system is in restricted-access stakeholder validation.

  • 5 planning layers
  • 2 utility briefings
  • ACEF 2025

Multi-agent planning · Demand forecasting · Geospatial analytics · FastAPI · deck.gl

system trace2025–26
01Grid
02Traffic
03Forecast
04Capacity
05Map
Public-safe planning system map

2024–25

Lead architect

Urban air-quality sensing network

ESP8266 sensor nodes publish through AWS IoT Core into DynamoDB. Lambda processing converts readings into EPA AQI bands, and a hosted dashboard renders IDW-interpolated heatmaps across the monitored area.

A disabled stream trigger produced no dramatic exception; it simply stopped forwarding data. By the time the gap became visible, months of readings were missing. I repaired the trigger path, added ingestion checks and changed interpolation so absent observations could not masquerade as a continuous surface. The climate programme spans eight partner universities.

  • 8 university partners
  • 4-stage cloud path
  • Months-long failure diagnosed

ESP8266 · AWS IoT Core · DynamoDB · Lambda · IDW interpolation

system trace2024–25
01ESP8266
02IoT Core
03DynamoDB
04Trigger
05Lambda
06AQI
silent boundarydisabled stream trigger · months without an alarm
Air-quality ingestion production trace

2025

Lead architect

Policy assistant for a national EV roadmap

An agentic RAG system over a 150-page national roadmap covering policy, finance and implementation. Retrieval returns section-level evidence; responses carry citations into the source document; evaluation flags unsupported claims before they enter a planning discussion.

The accuracy bar is unusually high because a fluent mistake can travel into a policy meeting and acquire authority it never earned.

  • 150 source pages
  • Section-level citations
  • 1 national roadmap

Agentic RAG · Hybrid retrieval · Claim evaluation · Citations · Guardrails

system trace2025
01150 pages
02Retrieve
03Rank
04Cite
05Evaluate
Roadmap retrieval and citation flow

2026

Designer and developer

Bilingual EV savings calculator

An Urdu and English progressive web app that models the economics of switching from combustion to electric against current fuel prices and levies. It was deployed as a kiosk and QR experience at a national EV expo, then adapted to capture adoption intent and waitlist interest.

A visitor can change the vehicle, distance and energy assumptions, then see where the calculation moves.

  • 2 languages
  • Kiosk + QR delivery
  • 1 public model

Progressive web app · Supabase · Cloudflare Pages · Scenario modelling

system trace2026
01گاڑی
02Distance
03Fuel
04Electric
05Savings
Monthly distance1,200 km2-language comparison
Interface reconstruction based on the public calculator

Public programme

Algorithms lead

EV retrofit engineering

Battery-sizing, charge/discharge-efficiency and thermal-management algorithms for combustion-to-electric drivetrain conversion. The calculations were used in a convertible motorcycle plug-in hybrid designed around a consumer-accessible price point in South Asia, with simulation assumptions aligned to Pakistani and EU green-mobility regulation.

The public campaign around the prototype reached roughly seven million views.

  • ≈7M campaign views
  • 2 regulatory contexts
  • 3 engineering models

Battery sizing · Efficiency modelling · Thermal management · Drivetrain simulation

system tracePublic programme
01Battery
02BMS
03Thermal
04Motor
05Range
Retrofit energy-flow diagram

Programme context

The clean-energy programme around this work includes more than 700 members from over 100 organisations. Its oversight committee draws from university energy institutes, policy centres and research laboratories. The programme is supported by an approximately USD 300,000 multi-year clean-energy award from an international climate foundation.