experience

click a role to see what i worked on.

performance boardmy most intensive project. the fastapi service behind it serves 48 endpoints across 12 domain modules, running split read and write postgresql pools on amazon relational database service (rds) with a replica lag guard, a redshift pool for warehouse queries, and a redis time-to-live (ttl) cache on amazon elastic container service (ecs). it is the one surface our 35+ team and executive leadership use instead of assembling fleet performance by hand. this is a company product, so access is by permission only.contact aarav for access
  • trained proximal policy optimization (ppo) reinforcement learning (rl) models in pytorch inside docker containers on ubuntu, running a pettingzoo multi-agent environment wrapping decision support system for agrotechnology transfer (dssat) crop simulations, shaping reward to trade yield against nitrogen cost so the policy learns application *timing*, not volume
  • lifted simulated corn yield 40% over the agronomist baseline at flat nitrogen across 800+ seasons, validating on held-out weather years to prove weather-adaptive timing over memorized seasons
  • ranked 25+ soil, weather, and management features by shapley additive explanations (shap) attribution against soil temperature and yield, rendering spatial density maps agronomists read directly and cutting analysis time 60%
  • integrated trained policies into the prescription pipeline, emitting georeferenced variable-rate field plans the robot fleet executes across 40 fields instead of notebook output
  • assisted in debugging row-following models, tracing perception failures that pulled robots off crop rows and reproducing them against logged field runs
  • scaled training on amazon web services (aws) sagemaker over 5 years of ontario weather and soil records joined to a year of internal nutrient and robot telemetry, running 200+ parallel hyperparameter sweeps
  • built python ingest pipelines with pandas and numpy over 10+ gigabytes of weather, soil, and yield data, normalizing per-field time series into training-ready tensors
  • automated an extract, transform, load (etl) pipeline on aws elastic compute cloud (ec2) under cron, extracting millions of postgresql records and loading 200,000+ rows daily into a redshift warehouse feeding power business intelligence (power bi) dashboards
  • modelled a dedicated warehouse_raw schema in amazon redshift around dashboard read patterns, with fact tables keyed on software version, field, device, and write time so version-over-version runtime comparisons resolve in 1 scan
  • replaced row-by-row inserts with amazon simple storage service (s3) staging plus redshift copy under an identity and access management (iam) role, cutting load times 80% and adding hourly per-table load logs so stale data surfaces instead of serving silently
  • launched 5 cron-driven slack bots on elastic compute cloud for sensor health, hourly fleet distance and reliability, weather off the open-meteo application programming interface (api), and a daily growing degree days (gdd) tracker predicting growth stages across 40 fields
  • stood up a company-wide notion ticketing system for 35+ staff, pulling daily, weekly, and monthly throughput metrics and advising stakeholders on delivery pace
  • automated sensor checks with chrome extensions and a short message service (sms) notification system, flagging drift and dropouts in under 5 minutes instead of manual fleet sweeps
  • shipped frontend portal work, carrying backend migrations through the user interface and closing 30+ bug and feature tickets on the internal product
  • migrated robot telemetry off polling onto websockets and zenoh publish/subscribe, a transport built for high-frequency fleet traffic, cutting message latency 50% in production
  • ran field operations across ontario corn fields, deploying and recovering robots in-row and collecting soil samples and crop observations that fed the ground-truth data behind the models
stackpythonpytorchppo / rlpettingzoodssatshapsagemakerdockerubuntuaws ec2ecsrdss3redshiftpostgresqlpandasnumpypower bizenohwebsocketsslack apiopen-meteonotion apichrome extensions
visit site
  • built a react and django dashboard covering 25+ projects on a team of 3, modelling the reporting domain in the django object-relational mapper (orm) behind normalized schemas and cached aggregates
  • designed idempotent sync jobs keyed on jira issue identifier and update timestamp, upserting rather than appending so 100% of duplicate rows vanished as upstream data drifted
  • replaced live jira application programming interface (api) reads with scheduled extract, transform, load (etl) pipelines, moving a per-project request fan-out on every page load into batched off-peak extraction into our own postgresql store
  • cut jira api calls 67% and page loads from 8 seconds to under 1 second by serving pre-materialized rows, leaving rate-limit headroom for every other team on the tenant
  • shipped the reporting surface for the commercial banking and payments technology quality assurance (qa) department, giving 50+ employees delivery status across every project in one view
  • ran demos and checkpoints with senior managers and directors across a 4-month build, keeping scope aligned with what leadership actually needed
  • engineered dagster etl pipelines over 100+ gigabytes of data as a software-defined asset graph, letting dagster resolve dependencies and run independent branches in parallel instead of a linear cron chain, saving 2+ days monthly
  • developed a python and graphql github activity tracker for 40+ offshore engineers, caching paginated commit queries in redis to automate velocity analytics across 500+ monthly commits
  • implemented jenkins and docker continuous integration (ci) pipelines and postgresql read replicas routing analytical reads off the primary, sustaining 99.9% uptime under 15,000+ concurrent queries
  • integrated scikit-learn anomaly detection behind a fastapi service scoring pipeline run metrics, surfacing outlier builds and cutting debugging time 40%
  • automated end-to-end quality assurance with selenium and playwright, converting manual regression passes into scripted cross-browser suites wired into continuous integration so every build self-verified, saving 15+ tester hours weekly
stackpythondjangoreacttypescriptfastapigraphqldagsterpostgresqlredisdockerjenkinsseleniumplaywrightci/cdscikit-learn
visit site
  • selected for cibc spark as 1 of 30 company-wide from 800+ interns, an accelerated development track reserved for top performers
  • coordinated the cloud@scale migration of 400+ applications to microsoft azure, tracking wave readiness and deliverables in azure devops against multi-year transformation goals and enterprise standards
  • automated triage reporting with excel visual basic for applications (vba) and power query, folding multi-source extracts into one refreshable model and cutting manual reporting 30%, saving 20+ staff hours weekly
  • managed planning for $3.5 million in initiatives using confluence and jira, tracking resource allocation and dependencies across 6 delivery teams
  • created real-time power business intelligence (power bi) dashboards over structured query language (sql) views of migration status, cutting leadership status-reporting turnaround from 3 days to same-day
  • authored weekly visual basic for applications and macro-driven excel reporting used by 10+ directors and senior directors as the standing source of truth for high-stakes portfolio decisions
stackazureazure devopssqlpower bipower queryexcel vbajiraconfluence
visit site

02 / 05