01
Preserve behavior before changing surfaces
Most of my strongest work is modernization work: changing protocols, tooling, and runtime paths while keeping business behavior stable enough for production teams to trust.
Software Developer / Backend & Platform / Gurgaon, India
Hi, I’m Ashutosh Modi. I build reliable backend platforms and developer tools.
I work on the careful side of software: modernizing long-lived C/C++ systems, building parser- and compiler-driven tooling, and shaping AI-assisted workflows that help engineers move faster without losing production judgment.
Download resume (PDF)
Read the Sandarbha case study
LinkedIn (opens in a new tab)
GitHub (opens in a new tab)
ashutosh94g@gmail.com
Open to backend, platform and developer-productivity roles · India / EU / Remote
$ sandarbha impact LEDGER_ENTRYsnapshot a41f…c9e2 sealed ✓writers 4+ 2 files unreadbatch jobs 3 exactservices 2 exactconflicts 1 reported, not resolved $ sandarbha conflictsLEDGER_ENTRY.STATUS 2 sources disagree no winner picked4+
Years building telecom backends
Amdocs · Gurgaon
2,800+
Automated tests in Sandarbha
Rust + UI, one release gate
500+
Legacy scripts auto-migrated
Lex/Yacc converter
4-5
Engineers, day-to-day execution
Review + unblocking
Tenure: 4+ years.
01 / About
Software developer with a systems-and-tooling bias.
I am a software developer at Amdocs, working on backend systems where reliability, migration safety, and developer productivity matter more than novelty. Most of my work sits at the boundary between old and new: C / C++ services moving toward typed RPC, legacy reporting moving into Python, and large legacy codebases becoming something engineers and AI agents can query with evidence instead of guesswork.
The thread through my projects is simple: I like turning repeated human work into checked, repeatable systems. That has meant Lex/Yacc compilers for migration and quality checks, Sandarbha, a Rust code-intelligence engine that answers questions about legacy code and says so when it cannot, shell and editor automation for day-to-day work, and Cursor knowledge bases that help agents respect the shape of a large codebase instead of guessing.
Since December 2025, I have also driven day-to-day technical execution for a 4–5 engineer team in an informal capacity: reviewing code, unblocking implementation details, and acting as the engineering manager’s primary technical interface. I am aiming for teams where that mix of careful backend work, automation, and practical AI tooling is useful.
02 / How I work
Careful with production, impatient with avoidable toil.
The projects below are the evidence. This is the pattern behind them: I like systems that are explicit, automatable, and understandable enough that the next engineer can safely keep moving.
02
Turn repeated work into checked systems
If a task repeats, I look for the structure underneath it. That is why parser-driven tools, generated code, build-time checks, and shell automation show up repeatedly in my work.
03
Use AI as an engineering system, not a shortcut
I care less about prompt tricks and more about evidence: giving agents source-backed knowledge, workflow rules and review loops, and building tools that say “I can’t tell” instead of guessing. Sandarbha is the sharpest version of that idea.
04
Keep execution visible
In team work, I naturally gravitate toward review, unblocking, and making decisions legible to peers and managers. Good engineering is easier when the state of the work is clear.
03 / Selected work
Selected engineering stories, not product pitches.
These are the projects I would talk through with another engineer: what was fragile, what I changed, how I reduced risk, and what the work taught me about building in large systems.
Flagship · case study
Active · Sep 2026 → present · Amdocs · on-prem
Sandarbha
A code-intelligence engine that refuses to guess.
Four evidence tiers (a data dictionary, a data-access layer, translated batch modules and a libclang compiler front end) are reconciled into one sealed, reproducible snapshot of a large legacy billing codebase. Engineers and AI agents query it through a CLI, an HTTP API, an MCP server and a web UI, and every answer either cites its source or says it cannot tell.
- When two sources disagree it reports the conflict. It never picks a winner.
- Rust workspace with a statically linked libclang and BLAKE3-verified SQLite snapshots.
- Mutation-verified tests, a reproducibility gate and checksum-verified standalone binaries.
- Evidence tiers
- 4
- dictionary, data layer, batch, compiler
- Automated tests
- 2,800+
- Rust + UI at one release gate
- Peak sync memory
- −34%
- rc.10 vs rc.6, measured once
- MCP first contact
- −83%
- 5-tool lens vs full catalogue
- Rust
- libclang
- SQLite
- MCP
- SolidJS
- GitLab CI/CD
Simplified architecture. Customer data, names and source are not shown on this site.
More engineering stories
04 / Stack
Skills grouped by the problems I solve.
01
Languages & data
- Rust
- C
- C++
- Python
- TypeScript
- SQL
- SQLite
- Oracle DB
- Linux/UNIX
02
Compilers & analysis
- libclang / Clang
- tree-sitter
- Lex/Yacc
- Static analysis
- Control-flow & data-flow graphs
- Parser-driven tooling
03
Distributed systems
- gRPC
- Protobuf
- HAProxy
- HashiCorp Consul
- HashiCorp Nomad
- Consul Template
- Apache Kafka
- Service discovery
- Load balancing
- High availability
04
Developer leverage
- Cursor
- MCP servers
- AI agent workflows
- Knowledge engineering
- Context design
- GitLab CI/CD
- SolidJS
- Git
- Linux shell
- Vim
- VS Code
- Tmux
- System design
- Data structures & algorithms
Code intelligence Compilers Static analysis Distributed systems AI agent workflows Telecom backend Service discovery Parser-driven tooling
05 / Experience
Four years, one company, deep work.
-
Sandarbha: code intelligence for legacy billing code
2026Designed and built a Rust platform that reconciles four evidence tiers, including a libclang-based compiler tier, into sealed, reproducible snapshots, then serves them through one query layer as a CLI, HTTP API, MCP server and web UI. Its rule: every answer cites its source or says it cannot tell. Read the case study →
-
Tuxedo → gRPC migration framework
2025Designed and shipped a generic protocol shim translating Tuxedo’s flattened buffers into strongly-typed Protobuf with zero changes to business code. Validated as a production-service POC and built for incremental rollout across long-lived telecom backends.
-
Day-to-day technical execution
LeadDrove technical execution for a 4–5 engineer team on a Tier-1 platform-modernization program since Dec 2025 — code review, unblocking, and the engineering manager’s primary technical interface in an informal capacity.
-
WebFOCUS replacement
PlatformReplaced TIBCO WebFOCUS reporting with an in-house Python library plus a Lex/Yacc auto-converter for 500+ legacy scripts, reducing dependence on licensed reporting tooling.
-
C/C++ stewardship
QualityHardened a long-lived C/C++ codebase: guard-clause refactors reduced potential vulnerabilities by 30%, a build-time prototype generator catches 80% of would-be production defects, and 95% of legacy compile-time warnings were resolved.
-
Internal AI development platform
AIOwned company-wide rollout of Cursor as the AI development platform — team knowledge base, agent skills, workflow rules, plus reusable Vim / VS Code snippets and shell automation. Cut team toolchain friction by 50%+.
06 / Contact
Let’s build systems that last.
Open to backend, platform, and developer-productivity roles at teams that care about reliability and engineering leverage — India, EU, or remote. Best reached over email; LinkedIn works too.