Juliu Pantia
AI Engineer
16+ years exp
JavaScript
SQL
Python
Bash
Experimentation & Evaluation: 4/10
Active 12 days ago
+7 (932) 6297883 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
LLM systems engineer (senior-level) specializing in agentic workflows and deterministic guardrails for honest outputs. The strongest proven skill is building robust, test-covered pipeline logic and deterministic honesty gates, evidenced by the audit rules and comprehensive FSM and audit tests (core/auditor_rules.py and tests/test_fsm.py). There is little or no evidence of custom model training, GPU/quantization work, or large-scale MLOps (serving/drift monitoring) in the human-authored code.
Technical skills
JavaScript
SQL• Senior • 16y+
Python• Middle
Bash• Junior
Python
Celery
pre-commit
Asyncio
FastAPI
Pydantic
SQLAlchemy
Databases
pgvector
PostgreSQL
Redis
AI/ML
Claude Code
ElevenLabs
Transformers
Anthropic SDK
Pandas
NumPy
Claude
LLM
Prompt Engineering
RAG
Sentence-Transformers
DevOps
Rest API
CI/CD
Docker
GitHub Actions
Grafana
Nginx
Prometheus
QA
Playwright
Pytest
Timeline
AI Analytics Engineer
•
Middle
Cipher Workshop
•
Full-Time
Built AI- and data-driven systems for business process automation. Designed multi-agent architectures combining LLMs with deterministic backend code, implemented RAG pipelines and validation/anti-hallucination checks, and delivered production services with end-to-end reliability practices. Developed backend data layers with FastAPI and PostgreSQL, and implemented CI/CD automation where agents generate and validate deployments. Implemented security features (SSO, OAuth, RBAC) and operational monitoring with Prometheus/Grafana and Telegram alerts.
Python
FastAPI
Asyncio
Pydantic
SQLAlchemy
PostgreSQL
pgvector
Claude
RAG
Sentence-Transformers
Redis
CI/CD
Docker
GitHub Actions
Nginx
Bash
Pytest
Grafana
Prometheus
Playwright
SQL
LLM
Prompt Engineering
Financial Analyst / System Analyst
•
Middle
NDA FinTech platform
•
Full-Time
Worked on investment accounting and financial modeling for a collective investments platform. Defined business methods (including profit allocation logic), translated requirements into IT specifications, and coordinated delivery with external IT specialists. Contributed to investor reporting and automation of primary documents using OCR and macros, and supported KPI and process design.
Bank Analyst (financial models and reporting automation)
•
Middle
Sber
•
Full-Time
Produced financial models and analytics for payroll-related business lines across multiple regions, including market and retention analysis. Built planning and management reporting models and standardized metrics for downstream reporting in BI tools. Worked with internal IT to convert business requirements into technical tasks, and accepted deliverables involving database and automation tooling.
SQLsince 2010
Tyumen State Academy of World Economy, Management and Law
Bachelor's Degree •
World Economy
Middle AI/ML Engineer
Confidence: Medium LLM Engineer
LLM systems engineer (senior-level) specializing in agentic workflows and deterministic guardrails for honest outputs. The strongest proven skill is building robust, test-covered pipeline logic and deterministic honesty gates, evidenced by the audit rules and comprehensive FSM and audit tests (core/auditor_rules.py and tests/test_fsm.py). There is little or no evidence of custom model training, GPU/quantization work, or large-scale MLOps (serving/drift monitoring) in the human-authored code.
Model Architecture & Training
How well models are designed and trained
Not evidenced in public code
Data Pipeline & Feature Engineering
3/10
How data is prepared for models
Concrete deterministic validation and payload/schema handling; evidence of careful data grounding and safe defaults but not large-scale ETL or feature-engineering pipelines.
Evidence
band-of-agents/core/auditor_rules.py: audit function that enforces grounded evidence and scoring rules
band-of-agents/core/payloads.py: Pydantic models for Candidate, Evidence, Verdict and typed payload validation
Experimentation & Evaluation
4/10
How results are measured and tested
Meaningful test coverage and focused edge-case tests that exercise the pipeline, recovery and honesty-gate behaviours; tests are deterministic and cover many failure modes.
Evidence
band-of-agents/tests/test_auditor_rules.py: unit tests covering grounding and blocking logic
band-of-agents/tests/test_fsm.py: tests for FSM progression, idempotency and corrupt-snapshot recovery
MLOps & Deployment
3/10
How models are shipped to production
Operational concerns are addressed - durable snapshots, restart recovery patterns and explicit run/supervisor code exist - but full model-serving or drift-monitoring pipelines are not present in the human-authored files.
Evidence
band-of-agents/demo_supervisor.py: process lifecycle management and graceful shutdown/startup handling
band-of-agents/core/fsm.py: explicit FSM design that supports durable state and application-level semantics
Computational Efficiency
1/10
How efficiently computing resources are used
Little to no explicit computational-efficiency engineering (no GPU/quantization/batching optimizations) in the human-authored artifacts.
Evidence
band-of-agents/room_io.py: async send/poll patterns show runtime-awareness but not low-level performance optimization
Research Depth & Innovation
2/10
Depth of research and new ideas
Some thoughtful deterministic guardrails and test-driven red-team fixtures show attention to honest model behavior, but there is no evidence of novel research-level model or algorithm development.
Evidence
band-of-agents/core/auditor_rules.py: deterministic honesty gate logic
band-of-agents/core/frozen_verdicts.py: red-team fixtures and reproducible blocked verdicts
Expertise
AI Agents & Agentic Workflows• Middle
Industries
Energy & Utilities• Middle
Financial Services• Middle
Sales & Marketing• Middle
Technologies
SQL• Senior • 16y+
PostgreSQL
Redis
Rest API
Claude
pgvector
Claude Code
Sentence-Transformers
SQLAlchemy
GitHub Actions
FastAPI
Prometheus
ElevenLabs
Prompt Engineering
CI/CD
Transformers
Docker
Nginx
Grafana
Celery
pre-commit
Recommendations
- Develop resilient multi-agent orchestration and guardrail layers that require determinism and idempotency (agent workflows, auditing, FSM-driven pipelines).
- Implement trustworthy RAG ingestion and deterministic grounding pipelines that trace facts to sources and reject unbacked claims.
- Build out production MLOps for inference: monitored serving, drift detection, and lightweight cost/latency budgets for LLM calls.
- Own data-validation and deterministic report generation for domain-specific workflows (finance or industrial equipment) where auditability is required.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Data Scientist
Confidence: High Data Engineer
A Senior-level data engineer specializing in deterministic, production-oriented pipelines and honest RAG/agent orchestration. The strongest proven skill is designing robust, idempotent async pipelines with durable state and deterministic anti-hallucination gates, demonstrated by the Planner FSM recovery/persist logic and the grounding/validation code (planner._persist/restore and groundtruth.extract/validate). There is limited evidence of statistical modeling, large-scale distributed data platforms, or advanced ML model training in the public code.
Statistical Rigor
4/10
Correct use of statistics
Solid deterministic checks and numeric guardrails are present, but formal statistical inference, hypothesis testing, uncertainty quantification and causal analysis are limited.
Evidence
churn-revenue-reporting-agent/src/agent_report.py:check_report_numbers
churn-revenue-reporting-agent/tests/test_metrics.py:test_month1_has_zero_churn
Data Wrangling & Cleaning
8/10
Preparing and cleaning data
Strong evidence of careful data ingestion, grounding, validation and safe defaults; deterministic grounding at ingest and many validation tests reduce leakage and fabrication risk.
Evidence
groundtruth-rag/src/groundtruth/sources/extract.py:is_grounded / build_records
groundtruth-rag/src/groundtruth/validate.py:validate_answer
churn-revenue-reporting-agent/src/generate_data.py:generate
Exploratory Analysis & Visualization
6/10
Exploring and visualizing data
Thoughtful demo rendering and deterministic report generation with narrative sections and custom visual helpers; visuals are purposeful and paired with textual interpretation.
Evidence
churn-revenue-reporting-agent/docs/onepagers/build.py:area_chart / bar_chart
groundtruth-rag/src/groundtruth/render.py:render_demo
Predictive Modeling
2/10
Building models that predict
Little evidence of predictive-model engineering or model tuning; work focuses on deterministic pipelines, metrics computation and optional LLM orchestration rather than training predictive models.
Evidence
churn-revenue-reporting-agent/src/agent_report.py:_agentic_report (LLM orchestration, not model training)
churn-revenue-reporting-agent/src/generate_data.py:generate (synthetic data generation, not modeling)
Business Insight & Impact
6/10
Turning analysis into business value
Analytical outputs are framed for business impact: deterministic takeaways, named actions and cost-relevant checks are present and tied to metrics.
Evidence
churn-revenue-reporting-agent/src/agent_report.py:compute_highlights and template_report takeaways
churn-revenue-reporting-agent/README.md (business brief intent and one-pager outputs)
Reproducibility & Notebook Hygiene
6/10
Clean, repeatable analysis
Good reproducibility practices: fixed seeds, tests, CI badges, atomic snapshotting and deterministic guardrails; environment pinning and data-versioning tooling are not present though.
Evidence
agents/planner.py:_persist and restore (atomic tmp write, fsync, quarantine on corrupt snapshot)
churn-revenue-reporting-agent/src/config.py:SEED and deterministic parameters
Expertise
Analytics• Middle
Streaming• Middle
Industries
Energy & Utilities• Middle
Financial Services• Middle
Technologies
Python• Middle
Anthropic SDK
Pandas
NumPy
LLM
RAG
Asyncio
Pydantic
Recommendations
- Develop production-grade observability and operational playbooks (metrics, tracing, alerting) around the async agent loops and persistence paths.
- Package the deterministic validation and grounding gate as a reusable library with clear API and integration tests for other RAG projects.
- Extend analytics work into formal statistical analysis and uncertainty quantification (confidence intervals, sensitivity checks) for higher-stakes business decisions.
- Harden deployment (container images, pinned dependencies, infra IaC) and add data-versioning (DVC or equivalent) for fully reproducible releases.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Junior Backend Developer
Confidence: Medium API Engineer
Backend engineer (middle-level) focused on building orchestrated agent-style services and deterministic validation gates. The strongest proven skill is deterministic anti-hallucination validation implemented in groundtruth.validate.py which enforces that every claimed fact or number is backed by cataloged source facts. There is little public evidence of production DB migrations, large-scale distributed-system design, or mature observability and operational playbooks.
API Design
4/10
How well APIs are designed
Conventional, well-typed API message shapes and an event/room client pattern are present with consistent stage/event contracts, but no explicit versioning or idempotency key policy beyond per-event keys is documented.
Evidence
band-of-agents/room_io.py: RoomClient / Outbound / wrap/parse message codec and send/poll lifecycle
band-of-agents/core/payloads.py: StageEvent, Verdict and other pydantic-typed payloads defining the API contract
Data Layer & Database
4/10
Working with databases
Clear ownership of data shape and deterministic validation logic; schema dataclasses and explicit validation/gates exist, but no migration history or explicit transactional boundaries for a database are present.
Evidence
groundtruth/schema.py: Record dataclass, record_to_dict / record_from_dict for canonical shapes
groundtruth/validate.py: ValidationError and validate_answer deterministic gate enforcing data integrity
Scalability & Performance
2/10
Handling load and speed
Some vector/array work and in-memory indexing appear (numpy usage), but no measured performance tuning, caching strategy, or production queueing patterns are demonstrated.
Evidence
groundtruth/index.py: tokenization/embedding helpers and use of numpy for matrices
groundtruth/retrieve.py: weighted lexical/vector retrieve logic
System Architecture
3/10
Overall system structure
Modular service-like separation (agents, planner, auditor, judge) and an FSM concept are evident, though service boundaries are small-scale and there is little evidence of operational contracts or multi-service deployment design.
Evidence
band-of-agents/core/fsm.py: PlannerFSM and Stage enum defining the state machine boundaries
band-of-agents/config.py: centralized agent/load_room_id configuration and env handling
Security & Auth
2/10
Protecting data and access
Basic secrets handling by expecting env variables and load_dotenv is used; some input validation exists, but there is little evidence of token lifecycle, revocation, or thorough dependency/security scanning.
Evidence
band-of-agents/agents/judge.py: build_agent checks ANTHROPIC_API_KEY environment variable
config.py (band-of-agents): dotenv usage and centralized config for URLs and keys
Reliability & Observability
3/10
Stability and monitoring
Good unit and integration tests around validation and FSM behavior increase reliability; some transient-failure handling and snapshot/quarantine ideas are present, but full observability, retry-with-backoff, and circuit-breaker wiring are not visible.
Evidence
groundtruth/tests/test_validate.py: comprehensive tests for validation gate and edge cases
band-of-agents/tests/test_loop_send_failure.py: simulated transient send failure and self-healing behavior
Expertise
Backend AI & LLM• Junior
Python• Junior
Industries
Energy & Utilities• Middle
Financial Services• Middle
Recommendations
- Develop API-driven RAG services that integrate the deterministic validation gate end-to-end (ingest -> index -> retrieve -> validate) using the existing groundtruth modules.
- Build small, test-first backend services around the agent workflows (Planner/Auditor/Judge) and add explicit inter-service contracts, versioning and API docs for consumers.
- Harden operational concerns: add explicit retry-with-backoff policies, timeouts, structured correlation ids in logs, and a basic metrics/alerting integration.
- If persistent storage is needed, add a migration history and explicit transaction boundaries (and tests) rather than relying on ad-hoc file snapshots.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: