Syed Ahnaf Raza syahra712

I build RAG systems, LLM agents and data pipelines, and I fix bugs in the open-source projects they're built on. When something breaks in a library I depend on, I read the source, reproduce the bug and send the fix upstream.

Right now

🔧 Contributing across the LLM-agent stack, with PRs open in litellm, crewAI, dspy, smolagents, mem0 and markitdown
🔤 Building urdukit, NLP tooling for Urdu, a language that 230M people speak and few libraries support
💼 PMO @ VentureDive · Software Engineering @ Bahria University

🔧 Merged upstream

Latest merge: Accept NumPy integers when indexing Grid, OneDGrid and PeriodicGrid in theochem/grid · 2026-11-03

Project	⭐	Area	What I fixed
docling-project/docling	`70k`	document AI · VLMs	Coerce MLX bfloat16 logprobs to float in VlmPredictionToken
BerriAI/litellm	`63k`	LLM gateway · billing	Don't retire a completed batch from cost recovery while output_file_id is lagging
nltk/nltk	`15k`	machine learning	Fix bare raise outside except block in sem.util.root_semrep Fix NameError in UndirectedTypeRaise.can_combine
sshuttle/sshuttle	`14k`	networking · firewalls	pf exclusions ignored under a wider include since rules became quick Skip hostwatch when the server has no os.fork Keep the tunnel up when a DNS or UDP reply cannot be sent Drop the zone id of a link-local nameserver from nft rules
biopython/biopython	`5.3k`	bioinformatics	Fix unformatted placeholder in double complement error message
tox-dev/tox	`4k`	Python test tooling	Keep the packaging backend captured with --no-capture Keep --no-index when an index URL follows it
pyro-ppl/numpyro	`2.8k`	bayesian inference	Parse JAX_ENABLE_X64 as a boolean string in enable_x64
aboutcode-org/scancode-toolkit	`2.7k`	license compliance	Cap pluggy below 1.7.0
neuropsychology/NeuroKit	`2.4k`	biosignals	Hrv_rsa(): compute RSA_PorgesBohrer on heart period, not heart rate Signal_formatpeaks(): drop every nan index together with its value
ceph/ceph-ansible	`1.8k`	ansible	Ceph_pool: fix arguments passed to list_pools()
pybamm-team/PyBaMM	`1.7k`	batteries	Fix kinetics `__all__` listing inverse kinetics submodules
fossasia/voxbento	`1.5k`	web API · access control	Scope admin transcript routes to the requested event
fossasia/veditor	`1.5k`		Show app INFO logs so the dev verification link is visible
mpmath/mpmath	`1.2k`	arbitrary precision	Flip rounding direction when negating in mpf_gamma(type=3) (fixes iv.loggamma) Respect directed rounding in mpf_atan2() (fixes iv.atan2 enclosures) Call lower_gamma() directly in the _gamma3() fallback (fixes #467)
pytroll/satpy	`1.2k`	closember	Fix debug log message for repeated record classes in eps_l1b reader
apache/iceberg-python	`1.2k`	apache	Preserve null maps when projecting a map of structs
SciTools/iris	`730`	data analysis	Show the offending object in CubeList non-cube error
pymc-devs/pytensor	`649`	ai	Index 0-d MakeVector inputs instead of calling .item() in numba
mozilla/bugbug	`576`	agentic ai	Keep full comment text when rendering Phabricator revisions
GoogleCloudPlatform/professional-services-data-validator	`535`		Convert Impala connection options given as strings to their real types Use UTF-8 for stdout so non-ASCII results can be printed on Windows Do not split --filters on PostgreSQL style :: type casts
readthedocs/sphinx-autoapi	`497`		Fix recursion when a nested class inherits from its enclosing class Fix method signatures when self is positional-only or absent
fortran-lang/fortls	`348`	fortls	Handle mixed quotes in find_paren_match
wradlib/wradlib	`323`	hacktoberfest	Show filename in read_gamic_hdf5 error for non-GAMIC files
gammapy/gammapy	`294`	astronomy	Fix missing f-prefix in two error messages
openSUSE/opi	`289`	cli	Fetch file:// and $releasever gpgkeys correctly when asking to keep a key
scikit-hep/uproot5	`274`	analysis	Handle negative and out-of-range indexes in TAxis.getitem
StingraySoftware/stingray	`227`	astronomy	Fix gti all entry and accept list GTIs in get_btis
getmail6/getmail6	`170`		#277: malformed POP3 UIDL/LIST aborts only the affected account #281: clear IMAP mailbox state when EXPUNGE fails in close_mailbox
scikit-hep/hist	`140`	histogram	Raise for missing exact names in wildcard category lists Accept NumPy integer arrays in from_columns
openfoodfacts/open-prices	`110`	crowdsourcing	API: accept `location_id` when updating a proof
theochem/grid	`102`	clenshaw curtis	Return an empty LocalGrid when no point is within the radius Accept NumPy integers when indexing Grid, OneDGrid and PeriodicGrid
pypa/distlib	`66`	Python packaging	Use the Python-2.0.1 SPDX license identifier
ga4gh/vrs-python	`65`	bioinformatics	Reject trailing input in beacon, gnomAD and SPDI translators
sugarlabs/www-v2	`62`	hacktoberfest	Redirect /music-blocks to /musicblocks
scipy/xsf	`43`		Smirnovp: avoid int overflow in derivative coefficient for large n
pyccel/sympde	`25`	finite elements	Keep the physical dimension of mappings in Domain.export / Domain.from_file
arviz-devs/arviz-stats	`18`		Stabilize stacking weights against overflow in exp of pointwise ELPD Fit the left Pareto tail correctly for unsigned integer draws
SciTools/ncdata	`16`		Replace deprecated licence config in pyproject.toml
konflux-ci/rpm-lockfile-prototype	`16`		Handle registry port numbers in image specs Add --tls-verify option for skopeo calls
fossology/fossology-python	`15`	automation	Don't send "None" as the folder description
plone/plone.event	`5`		Pass re.MULTILINE to re.sub as flags, not as count
apache/solr-orbit	`3`	apache	Build the Docker image on Python 3.13, inside requires-python Page through collections with SolrClient.search in create-workload Restore metrics.metrics_store_class so solr-orbit stop works again Create ClusterEnvironmentInfo so --telemetry cluster-environment-info works Add a make licenses target that generates THIRD_PARTY.md

🕐 In review: 84 PRs across 70 projects, including markitdown, mem0, litellm, crewAI, agno, dspy, smolagents…

Project	⭐	Fix
microsoft/markitdown	`194k`	Refactor ImageConverter to reuse llm_caption helper
mem0ai/mem0	`68k`	Dispatch add_embeddings on actual signature
BerriAI/litellm	`63k`	Route GPT-5.4+ function-tool calls through the Responses API
crewAIInc/crewAI	`60k`	Opt-in deferred tool loading via Responses tool search
agno-agi/agno	`43k`	Chain exception causes in tool, model, and vectordb error paths
stanfordnlp/dspy	`39k`	Add sort_keys=True to tool call/result JSON serialization
huggingface/smolagents	`30k`	Accept AgentImage/AgentAudio state values for image/audio tool inputs
ArduPilot/ardupilot	`16k`	Tools: param_unpack.py: support param.pck files with defaults
pyro-ppl/pyro	`9.1k`	Restore deep_getattr in pyro.infer.autoguide.utils
arrow-py/arrow	`9.1k`	Use calendar months for humanize month and quarter granularity
marshmallow-code/marshmallow	`7.2k`	Reject leading/trailing hyphen in the last domain label in validate.Email
python-pendulum/pendulum	`6.7k`	Fix precise_diff counting partial months as full months

_{See every open PR →}

_{This section updates itself every 6 hours from the GitHub API. See scripts/update.py.}

🚀 Things I've built

🔤 urdukit

A zero-dependency Urdu NLP toolkit for normalization, tokenization, stopwords, stemming and Roman-Urdu transliteration. pip install and go: no TensorFlow and no model downloads.

Python NLP low-resource languages

🛠️ HizmatAI

A two-sided agentic booking platform for Pakistan's informal economy. A 6-agent Google ADK pipeline on Gemini connects customers with home-service providers in Urdu, Roman-Urdu and English.

multi-agent Flutter FastAPI React

🕸️ Graph-Augmented Research Agent

A research agent that goes beyond plain RAG: it combines embedding similarity with co-citation and bibliographic coupling for literature search, gap analysis, plagiarism detection and citation formatting.

RAG citation graphs LLM agents

🌫️ AQI forecasting pipeline

An end-to-end MLOps pipeline. Scheduled ingestion feeds a Hopsworks feature store, models retrain automatically, and a live dashboard shows the forecasts, all running on GitHub Actions.

MLOps feature store time series

🫥 hiddenink

Find the ink you can't see. It detects and cleans invisible Unicode, homoglyphs and container metadata, and decides for each occurrence whether the character is a hidden mark or real content.

Python text security zero-dep

🌳 yaadain

A living family tree for Urdu-speaking elders with memory loss. It's an offline-first Flutter app built for reminiscence and simulated-presence therapy.

Flutter Dart health

_{More: BridgeSign (real-time ASL recognition) · EcoViz (emissions dashboards and model benchmark) · TumorVision (MRI classification with Grad-CAM) · AES-RSA (hybrid encryption service)}

🎖️ Experience & credentials


💼 Experience	PMO @ VentureDive · Technical APM @ CodeAvenue · RAG / LLM systems @ Sybrid · ML pipelines @ 10Pearls
🏆 Awards	Regional Winner, National Agentic AI Hackathon
📜 Certs	Oracle Generative AI Professional · Stanford ML Specialization · Microsoft PL-300 · Google Project Management
🌱 Leadership	Hult Prize Campus Director · VP, Bahria University Development Society

🛠️ Stack


AI / ML	PyTorch · TensorFlow · Hugging Face · LangChain · Google ADK · spaCy · RAG · LLM agents
Languages	Python · TypeScript · JavaScript · Dart · SQL
Build & ship	FastAPI · React · Next.js · Flutter · Docker · MongoDB · GitHub Actions · n8n

_{Open to AI/ML engineering and applied AI project management roles · If you maintain a Python project and have a gnarly bug, I'd like to hear about it.}

Navigation Menu