Software EngineerAI infrastructure and systems researchEdinburgh, UK
VidhunVijayakumar Suja
I build complex systems and then take them apart to see what actually matters. Lately that means an LLM serving engine written from scratch, a controller that tries to tune it while it runs, and an SMT-based checker for weak memory models.
I care about measuring carefully and about saying exactly how far a result goes. MSc Software Engineering from Heriot-Watt, with three manuscripts under review.
- 01LLM inference and serving
- 02Adaptive runtime control
- 03Weak memory and concurrency
- 04SMT and Isabelle/HOL
- 05Distributed backends
- Focus
- LLM serving, adaptive runtime control, weak memory verification
- Education
- MSc Software Engineering, Heriot-Watt. Dissertation graded Exceptional (78/100)
- Research
- Three manuscripts under review: MLSys, TACAS and FSEN 2027
- Experience
- Research software engineering intern, retrieval and RAG, Dec 2023 to Aug 2024
From retrieval pipelines to memory models
I studied computer science in Kerala, then spent eight months as a research software engineering intern on a five-person team building retrieval systems. That is where I learned that the interesting failures live in the plumbing: a reranker that dominates p95 latency, or an ingestion job that duplicates records after a crash.
At Heriot-Watt I took concurrency and formal methods seriously for the first time. My dissertation became a tool for stepping through executions under WEAKEST, a memory model that is as permissive as possible while still ruling out values that appear out of thin air. That tool grew into WEV-SMT, an SMT-backed checker, and into two papers about what dependency analyses can and cannot know.
Alongside it I wanted to understand LLM serving from the inside, so I wrote an engine: Mini-vLLM. Then I built CARL, a controller meant to tune it online, and spent longer auditing CARL than building it. Several of my early numbers did not survive that audit. The paper that came out of it reports what did.
Measure first
Numbers here say where they came from: a log, a results file, a manuscript, or my own report.
State the scope
One GPU, one model and two knobs is exactly that, and I say so.
Make it recoverable
Checkpoints, hashes and regression tests for the boring parts.
Two research systems, and the rest
The first two are where most of my time went this year. Each case study has interactive figures, the measured results, and the parts that did not work.
Also featured
- 03Event-driven e-commerce backendSeven Spring Boot services, a choreographed saga over Kafka, and measured resilience behaviourPersonal project
- 04Verified autonomous intersectionTwo controllers, three vehicle types and six LTL properties, model checked in SPINCoursework
- 05Winter cabin diorama in WebGPUA real-time scene with WGSL Phong shading, an OBJ loader and simple physicsGroup coursework
Papers in review, and how they got there
Nothing below is accepted or published yet, and the labels say so. Each manuscript has its code public, and each states its scope: one model and one GPU for CARL, straight-line 32-bit code for the dependency theorems.
- Under reviewMLSys 2027 · Oct 2026
CARL: Comparing an Online Contextual-Bandit Reconfiguration System with Offline-Tuned Static Configurations of an LLM Serving Engine
V. Vijayakumar Suja
A paired comparison over ten seeds on a Tesla T4. At a saturated burst point the controller beat a non-learning rule by 3.76 tok/s but stayed below static configurations chosen offline for the same objective. At a sub-critical Poisson load it trailed both on end-of-run utility, and no run met the convergence diagnostic.
- Under reviewTACAS 2027, submission #4549 · Oct 2026
The Limits of Instruction-Local Dependency Summaries, with an Application to Thin-Air Reads
V. Vijayakumar Suja, M. Doko
Proves in Isabelle/HOL that no Boolean depends-or-not summary restricted to an instruction's read registers stays exact under composition, and that composed expression labels do. With an exact per-instruction SMT oracle, the prototype's closure catches all 66 single-instruction fake dependencies but none of the 72 written as chains, while keeping all 171 real ones.
- Under reviewFSEN 2027, submission 41 · Nov 2026
The Limits of Read-Restricted Dependency Summaries
V. Vijayakumar Suja
Characterises what an exact instruction-local summary must keep and when Cousot's composition is already exact. The empirical side is negative and reported as such: across 2,250 litmus tests and eight upstream concurrency code bases, no composed cancellation appeared in a store's data dependencies.
- DissertationMSc dissertation, Heriot-Watt University · Apr 2026
Visualising WEAKEST Executions
V. Vijayakumar Suja, supervised by Dr M. Doko
Design, implementation and evaluation of an interactive JavaFX tool for exploring executions under the WEAKEST memory model, including a user study with nine participants. Graded Exceptional (78/100).
Research timeline
Oct 2025
Ethics approval and project proposal
Proposed an interactive tool for exploring executions of concurrent programs and applied for ethics approval for a user study.
Apr 2026
Dissertation submitted
WEAKEST Execution Visualiser, with a five-model comparison view and a nine-person user study.
May 2026
Mini-vLLM engine work begins in earnest
Continuous batching, paged KV cache and HF parity tests, then prefix caching, chunked prefill and metrics.
Aug 2026
WEV-SMT reaches real corpora
Litmus parser for five architectures, validation over herd7 and Dat3M tests, fence and RMW encoding.
Sep 2026
CARL audit and repair
A claim ledger retracted results the repository's own artifacts contradicted. The controller and reward were repaired and re-run on a T4.
Oct 2026
Two submissions
TACAS 2027 on dependency summaries, and MLSys 2027 on CARL, each with three-campaign provenance and stated scope.
Nov 2026
FSEN 2027 submission
Read-restricted dependency summaries, with a negative result from real code reported alongside the theorems.
Everything else, filed properly
Coursework and side projects, described from their source code and reports. The archive has filters by category, technology and type.
MSc dissertation · 2026
WEAKEST Execution Visualiser
MSc dissertation: a JavaFX tool for stepping through concurrent executions
Java 21 · JavaFX · GraphStream · Maven
Coursework · 2025
Stall prevention system in SPARK
High-integrity Ada for an aircraft angle-of-attack monitor
SPARK · Ada · GNAT
Coursework · 2025
Document Tracker analytics
Views by country, reader profiles and an also-likes graph over Issuu logs
Python · Tkinter · matplotlib · Graphviz
Coursework · 2025
Simple web browser in C#
Raw HTTP, history, bookmarks and a homepage in Windows Forms
C# · .NET 8 · Windows Forms
Personal project · 2026
AI resume analyser
FastAPI and React app that scores a CV and checks it against a job description
Python · FastAPI · SQLAlchemy · React
Full archive
All 10 projects
Skills, each with a receipt
Pick an area, then a skill, to see which projects actually use it.
How a serving engine decides what runs on the GPU next, and how to tell whether a change helped.
Projects behind ai systems and inference
Mostly mediums, all in Java
A snapshot of my public LeetCode profile, taken on 18 November 2026. LeetCode has no stable public API, so these numbers are cached rather than live.
Solved
153
problems · Java for every one
- Easy
- 33
- of 976
- Medium
- 93
- of 2,148
- Hard
- 27
- of 991
Most practised topics
Problems solved per tag. A problem can carry several tags.
- Array68
- String37
- Dynamic Programming33
- Tree26
- Hash Table23
- Two Pointers23
- Linked List22
- Depth-First Search20
- Backtracking19
- Math19
- Breadth-First Search13
- Stack13
Where I have been
Until Jul 2023
Education
BTech, Computer Science
College of Engineering Muttathara, Kerala
Undergraduate computer science: data structures, operating systems, networks and databases.
Dec 2023 to Aug 2024
Experience
Research Software Engineer Intern, Retrieval and RAG
Kompetenzen Technologies, Thiruvananthapuram
Part of a five-person research team building document intelligence and retrieval systems.
- Hybrid retrieval with BM25, dense embeddings and cross-encoder reranking. nDCG@10 went from 0.58 to 0.68 on 500 held-out queries.
- Citation mapping and evidence-based abstention cut unsupported claims from 18% to 9% at 80% answer coverage.
- Bounded candidate sets, batched reranking and async handling took p95 retrieval latency from 1.2 s to 0.8 s at 20 users.
- Checkpointed, transactional ingestion with content-hash deduplication: zero duplicate records across 100 injected-failure tests.
Self-reported· figures from internal evaluation; no public artifactAwarded Jun 2026
Education
MSc Software Engineering
Heriot-Watt University, Edinburgh
Machine Learning, Concurrent Systems, Formal Methods, Distributed Computing, Rigorous Methods, Industrial Programming and 3D Graphics. Dissertation graded Exceptional, 78/100.
Jan 2026 to present
Research
Independent systems research
Edinburgh
Mini-vLLM and CARL on the inference side; WEV-SMT and two dependency-summary papers on the verification side. Three manuscripts under review.
08 Contact
If you work on serving, schedulers or correctness, I would like to hear from you.
I am looking for software engineering roles in AI infrastructure, inference and systems. I am based in Edinburgh with the right to work in the UK (UK Graduate Route visa, valid through October 2028).