Sepehr Vahedi
Software engineer
Backend, distributed systems, ML infrastructure

London, UK
Available from September 2027
sepehr.vahedi@gmail.com

§1 Abstract

I make production systems measurably faster, and then prove they still give the same answers.

Sepehr Vahedi

For 3.5 years I worked full-time at Tapsi, a ride-hailing and delivery app with more than ten million monthly users, alongside my BSc at Sharif University of Technology. I was promoted to Senior Software Engineer in January 2024. I owned services from design to code review and on-call.

I am now doing an MSc in Software Systems Engineering at UCL, building a patient-facing LLM agent part-time at PillSorted, and looking for a software engineering role in the UK from September 2027, on a Graduate visa with no sponsorship needed to start.

Away from work I am a Candidate Master on Codeforces (max rating 2075).

§2 Selected results

Five numbers, what changed to produce each one, and how I know it is real.

ResultWhat changedHow I know
54× throughput ETA model server at Tapsi: 70 → 3,800 requests/s per pod, and p50 latency at peak 866 → 62 ms. Removed a serialisation bottleneck in NVIDIA Triton and batched the sub-model calls. Outputs identical on 19,200 replayed requests before it shipped.
24 → 19% ETA error Took the delivery domain's first ML models to production on Triton, with accuracy gates in CI. Six-city shadow run, then a switchback A/B test (MAPE).
−45% infrastructure cost Rewrote the service behind every ride request from a Node.js monolith to Spring Boot and Project Reactor. Peak response time 200 → 80 ms. Zero downtime during the switch.
2× daily deliveries Built the event-sourced Kafka backend (1M+ events a day) that lets a customer run several delivery orders at once. Measured after the multi-delivery launch.
92.5% branch coverage TestCouncil, my BSc thesis: three different LLMs write unit tests for the same function and a synthesis step merges them. 135 functions; 36+ points of line coverage above the best single model, p < 0.001.
Requests per second per pod before and after: 70 and 3,800. Median latency at peak before and after: 866 and 62 milliseconds. beforeafter 70 req/s3,800 req/s per pod beforeafter 866 ms median at peak62 ms
Fig. 1. The ETA model server before and after, drawn to scale. Top: throughput. Bottom: median latency at peak.

§3 Experience

Nov 2026 – now

Software Engineer, part-time · PillSorted, NHS online pharmacy, Cambridge

  • Sole engineer on Ask PillSorted, an iOS and Android chat that answers patients' questions from their own pharmacy records.
  • Designed its LLM agent on Amazon Bedrock: tools over MCP, rule-based guardrails, and retrieval that cites NHS sources.

Dec 2025 – Sep 2026

Senior Software Engineer, Delivery and Marketplace · Tapsi, Tehran

  • Designed and owned the delivery-routing gRPC service (Kotlin, Spring WebFlux) behind every motorcycle ETA and distance estimate, with an ML → OSRM → Haversine fallback chain. Moved five services off the old Go estimator with no regressions.
  • Raised ETA model-server throughput 54× and cut p50 latency at peak from 866 ms to 62 ms (figure 1).
  • Took the delivery domain's first ML models to production on NVIDIA Triton (KServe v2 gRPC, CI/CD with accuracy gates): ETA error fell from 24% to 19%.
  • Led these projects with a data scientist, owned code review and on-call for the routing stack, and was the onboarding mentor for a new engineer.

Jan 2024 – Dec 2024

Senior Software Engineer · Tapsi

  • Built the event-sourced Kafka backend (1M+ events a day) for multi-delivery orders. The launch doubled daily deliveries.
  • Rewrote the ride-request service, which manages the lifecycle of every ride request, from Node.js to Spring Boot and Project Reactor with zero downtime: 45% lower infrastructure cost.

Feb 2022 – Dec 2023

Software Engineer · Tapsi, co-op then full-time alongside my BSc

  • Built the backend for a new business line, Iran's first school-transport platform (2,000+ drivers, 12,000+ students), as one of two backend engineers on a six-week deadline.
  • Cut GPS-service load 67% with adaptive driver GPS intervals, caching and filtering, and rewrote reverse geocoding in Go.
  • Built a mesh-level rate limiter on my own initiative. Every service exposed to external clients adopted it (figure 3).

Feb 2020 – Feb 2021

Co-founder and product designer · RabinCash, crypto exchange, Tehran

  • Owned product and UX design for an exchange with 20+ trading pairs, and led the two-person front-end team.

§4 Research

2025 – 2026

TestCouncil: multi-agent LLM unit-test generation · BSc thesis, Software Quality Research Lab, Sharif

Qwen, DeepSeek and Gemini write tests for the same function in parallel; an LLM synthesis step merges them and removes about 85% of the tests as redundant. The finding I did not expect: giving the models personas made no significant difference. Using different models did.

Average line coverage: Qwen alone 43.0%, DeepSeek alone 38.6%, Gemini alone 49.0%, council without roles 84.9%, council with roles 84.6%. Qwen aloneDeepSeek alone Gemini alonecouncil council + roles 050100% 43.038.649.0 84.984.6
Fig. 2. Average line coverage over 135 functions from HumanEval, MBPP and TheAlgorithms. Supervisor: Prof. Seyed-Hassan Mirian-Hosseinabadi.

2025 – 2026

LLM issue-title generation · Intelligent Software Engineering Lab, Sharif

A learned selection step decides which parts of a GitHub issue an LLM actually needs in order to write a good title: 37.3% fewer input tokens, within 0.004 BERTScore of the GPT-4 baseline, on a dataset of 2.6M+ issues. Supervisor: Prof. Abbas Heydarnoori.

§5 Open source

Envoy · Lua · Istio

local-rate-limit

Per-user, per-API rate limiting at the proxy, so no service has to implement it. A small Lua filter works out who is calling; Envoy's token buckets do the rest. A clean re-implementation of the pattern I first built at Tapsi, with integration tests against a real Envoy.

5 of 5 tokens
  1. Click faster than one request every two seconds.
Fig. 3 (interactive). A token bucket of 5 per 10 seconds: a burst of five, then one token back every two seconds. The same limit the repository's tests pin down.

Python · agent skill

grad-tracker

Every UK graduate role in one list, ranked against your CV, running on your own laptop. I replaced its first LLM scorer with a few hundred lines of rules: rank correlation 0.83 with the LLM's scores on a 290-role benchmark, in under a second, for nothing. It also installs as a skill for AI coding assistants, where the code keeps the scoring and the model only does the talking.

Python · LLMs

automatic-paper-critique

Four LLMs critique a paper and rank each other's points in shuffled order, to cancel position bias. A Borda count merges the rankings into one review.

All repositories, including a C-Minus compiler front end and a venue-booking platform on Spring Boot and Kafka.

§6 Also

Competitive programming
Codeforces Candidate Master, max 2075 (top 1.5%); Expert on an earlier account; 420+ problems. Iran National Olympiad in Informatics, top-50 finalist in 2018 and 2019. Third of 65 teams at the AUT programming contest.
Teaching
Taught informatics olympiad students at NODET; they won one gold and two silver national medals. Teaching assistant for Advanced Programming and for Scientific and Technical Presentation at Sharif.
Education
MSc Software Systems Engineering, University College London, 2026 – 2027. BSc Computer Engineering, Sharif University of Technology, 2020 – 2026: 18.25/20 over the final two years, six years in total because I worked full-time for most of them.
Languages
Kotlin, Java, Python, TypeScript, Go, C++, C#, Swift.
Backend
Spring Boot, Project Reactor, gRPC, REST, Kafka, event-driven microservices, JUnit, Mockito.
ML and LLM systems
NVIDIA Triton, KServe v2, ONNX, PyTorch, Amazon Bedrock, MCP, RAG, agents.
Infrastructure
Kubernetes, Docker, Helm, ArgoCD, GitLab CI/CD, Istio, Envoy, Terraform, AWS, Prometheus, Grafana.
Data
PostgreSQL, pgvector, MongoDB, Redis, Hazelcast, Uber H3, OSRM.
Spoken
English (IELTS 8.0), Persian (native), German (intermediate).

§7 Contact

Hiring for backend, infrastructure or ML-platform engineering from September 2027? I would like to hear about it.

CV: backend version, ML-platform version, or plain text.